Decisions With Delayed PK/PD Outcomes

Why delayed outcomes change an interim decision

At an interim analysis, concentrations may already be available for many participants while pharmacodynamic measurements or clinical endpoints are still being collected. The absence of a mature outcome does not mean that the participant had no response, nor does it mean that the available concentration data directly establish clinical benefit. It means the decision must account for what is known and what remains uncertain about outcomes that have not yet been observed.

A longitudinal PK/PD and clinical outcome model links dosing and concentration histories to evolving pharmacodynamic measures and, where appropriate, to clinical outcomes. At the interim, posterior prediction uses that model and the data currently available to describe possible later outcomes. The central task here is to translate those predictions into a prespecified adaptation decision: for example, continue a regimen, change a dose, pause enrollment, or stop a regimen.

Definition: Let \(D_t\) denote all trial data available at interim time \(t\), including observed doses, concentrations, covariates, observed portions of outcome trajectories, and the follow-up status of each participant. A delayed outcome is an outcome whose relevant measurement time has not yet occurred or whose measurement is not yet available at \(t\). A posterior predictive distribution is the distribution, conditional on \(D_t\), of an unobserved outcome or future outcome summary, integrating over uncertainty in the model parameters and latent participant quantities.

From interim data to a predictive decision quantity

The interim data should be represented according to when and how each measurement was obtained. For a participant with early concentration measurements but no mature clinical endpoint, the observed concentrations contribute information about exposure and model parameters. The unobserved endpoint is predicted from the fitted longitudinal model; it is not entered as a zero or treated as if it had already been measured.

A decision quantity should match the clinical question. It might be the probability that a regimen's mean clinical benefit exceeds a minimum worthwhile level, the probability that a future participant attains a pharmacodynamic target, or the probability that the toxicity rate exceeds an unacceptable level. These are different targets, so the decision rule must specify which one is being used, its threshold, and the action attached to that threshold.

For example, if \(\Psi_A\) is the true mean clinical response under regimen \(A\), a benefit rule might continue the regimen when \(P(\Psi_A \geq \psi_{\min}\mid D_t)\) exceeds a chosen cutoff. This is a posterior probability about a parameter or estimand, not a frequentist p-value. Alternatively, if the target is an individual future response \(Y_{\mathrm{new}}\), the rule may use \(P(Y_{\mathrm{new}}\geq y_{\min}\mid D_t)\). Individual outcome variability generally makes that predictive probability different from the probability about the mean.

Key distinction: A posterior distribution for a parameter describes uncertainty about a model quantity, such as \(\Psi_A\). A posterior predictive distribution describes uncertainty about an outcome that could be observed, including both uncertainty about model quantities and variability in outcomes conditional on those quantities. State which probability the adaptation rule uses.

Why posterior prediction includes two sources of uncertainty

Suppose \(\theta\) represents model parameters and latent quantities relevant to a future endpoint \(Y_{\mathrm{new}}\). The model first gives a distribution for \(Y_{\mathrm{new}}\) conditional on \(\theta\). The interim analysis then averages that distribution over the posterior distribution of \(\theta\) given \(D_t\). This integration is what allows observed concentration and partial trajectory data to inform predictions without pretending that uncertain parameters are known exactly.

p(Y_{\mathrm{new}}\mid D_t) = \int p(Y_{\mathrm{new}}\mid \theta,D_t)\,p(\theta\mid D_t)\,d\theta

The associated variance decomposition follows from the law of total variance. Add and subtract the conditional expectation \(E(Y_{\mathrm{new}}\mid\theta,D_t)\) around \(Y_{\mathrm{new}}\), and take expectations conditional on \(D_t\). The conditional residual has mean zero given \(\theta\) and \(D_t\), so its cross-product with the conditional mean contributes zero. Therefore:

\operatorname{Var}(Y_{\mathrm{new}}\mid D_t) = E\!\left[\operatorname{Var}(Y_{\mathrm{new}}\mid\theta,D_t)\mid D_t\right] + \operatorname{Var}\!\left(E[Y_{\mathrm{new}}\mid\theta,D_t]\mid D_t\right).

The first term is average outcome variability conditional on the model quantities; the second is uncertainty about the predicted mean as those quantities vary over their posterior distribution. Ignoring either term can make a prediction interval too narrow and can make a threshold-based adaptation rule appear more decisive than the data justify.

Worked Example: Predicting Whether Mean Benefit Clears a Target

Hypothetical scenario. At an interim, a fitted longitudinal model has used available concentration histories and partial outcome trajectories to estimate the posterior distribution of \(\Psi_A\), the mean change in a clinical score under regimen \(A\). Larger positive changes are better, and the minimum worthwhile mean change is 10 points. Suppose the model gives \(\Psi_A\mid D_t\sim N(10.8,1.8^2)\). For this calculation, assume the model and the normal posterior approximation are adequate for the interim data.

The probability that mean benefit reaches the target is:

P(\Psi_A\geq 10\mid D_t) = 1-\Phi\!\left(\frac{10-10.8}{1.8}\right) = \Phi(0.4444) \approx 0.672.

The posterior probability is approximately 0.672 that the regimen's mean score change is at least 10 points, given the interim data and model. If the prespecified rule requires this probability to exceed 0.80 to continue unchanged, this benefit criterion is not met. That is not proof that the regimen fails; the posterior probability is below the chosen decision cutoff.

The units remain important: \(\Psi_A\) and the 10-point target are both in score points. The probability itself has no units. This calculation concerns the mean benefit, not the probability that an individual participant improves by 10 points.

Pending outcomes and safety decisions

The same logic applies to delayed safety outcomes, but the model and target must reflect the safety endpoint and its observation window. If a toxicity can only be assessed after a specified follow-up period, a participant who has not completed that period is not automatically a nontoxic participant. The interim analysis can either wait for more follow-up or predict the eventual safety outcome under an appropriate model.

Worked Example: A Posterior Safety Rule With Mature Follow-Up

Hypothetical scenario. Among participants with completed safety windows, 7 of 10 experienced a prespecified toxicity. Assume a uniform \( \operatorname{Beta}(1,1) \) prior for the regimen's toxicity probability \(p\), and, conditional on \(p\), assume these outcomes are independent Bernoulli observations with a common probability. The posterior is \(p\mid D_t\sim\operatorname{Beta}(8,4)\). Suppose the rule is to pause the regimen if \(P(p>0.5\mid D_t)>0.80\).

For integer beta parameters, the beta-binomial identity gives \(P(p>0.5)\) as the lower-tail probability \(P(X\leq 7)\) for \(X\sim\operatorname{Binomial}(11,0.5)\). Thus:

P(p>0.5\mid D_t) = P(X\leq 7) = \frac{\binom{11}{0}+\binom{11}{1}+\cdots+\binom{11}{7}}{2^{11}} = \frac{1816}{2048} = 0.8867.

Since \(0.8867>0.80\), the rule calls for pausing the regimen. The calculation is conditional on the stated prior and binomial model; it does not establish that the toxicity probability is certainly above 0.5. If other participants have not completed their safety windows, this calculation describes the mature observations only. A decision using pending participants must additionally model their eventual outcomes rather than silently count them as nontoxic.

Predicting the completion of a pending cohort

A posterior predictive decision can also ask what is likely to happen among participants whose outcomes have not matured. This matters when the rule is framed in terms of an enrolled cohort's eventual results. The following example uses a simple binary outcome model to show the calculation. In a longitudinal pharmacometric analysis, the prediction would ordinarily condition on the participants' exposure histories, covariates, and available outcome trajectories as well.

Worked Example: Predicting Responses Among Participants Awaiting Assessment

Hypothetical scenario. Ten participants have mature binary response outcomes, with 6 responders. Use a \( \operatorname{Beta}(1,1) \) prior and assume that, conditional on the response probability \(p\), outcomes are exchangeable independent Bernoulli observations. The posterior is \(p\mid D_t\sim\operatorname{Beta}(7,5)\). Four more participants are enrolled but their responses are pending. The rule continues the regimen if the probability that at least 8 of all 14 participants eventually respond exceeds 0.70.

There are already 6 observed responses, so at least 2 of the 4 pending participants must respond. Let \(K\) be the number of responses among those 4. Integrating over the uncertain \(p\) gives a beta-binomial predictive distribution:

P(K=k\mid D_t) = \binom{4}{k} \frac{B(7+k,\,5+4-k)}{B(7,5)},\qquad k=0,1,2,3,4.

Substitution gives \(P(K=0\mid D_t)=0.05128\) and \(P(K=1\mid D_t)=0.1795\). Therefore:

P(K\geq 2\mid D_t) = 1-0.05128-0.1795 = 0.7692.

Because \(0.7692>0.70\), the specified rule continues the regimen. This probability is for at least 8 total responses among the 14 enrolled participants, not for the response probability \(p\) exceeding 0.5. The beta-binomial calculation allows for uncertainty in \(p\); simply substituting its posterior mean \(7/12\) and treating it as known would omit that parameter uncertainty.

Making the rule operational

A usable interim procedure needs more than a predictive model. Before the decision is made, the protocol or analysis plan should identify the decision time, the data cut, the predictive target, the model used to obtain it, the thresholds, and the actions attached to each region. For multiple regimens, the rule should also specify how predictions are compared and what happens if no regimen meets a criterion.

Define the interim data.
Record which doses, concentrations, covariates, partial trajectories, and outcome assessments are available, including the follow-up status of each participant.

Fit or update the longitudinal model.
Use the specified PK/PD and clinical outcome structure to update parameter and latent-quantity uncertainty using the interim data.

Compute the target prediction.
Integrate over posterior uncertainty to calculate the stated probability or predictive summary for benefit, safety, or eventual cohort outcomes.

Apply the prespecified action rule.
Compare the prediction with the decision thresholds and implement the action assigned to the resulting region; do not choose a threshold after seeing the interim result.

The assumptions must be credible at the interim. These include an adequate longitudinal model for the observed concentration and outcome trajectories, a defensible link between exposure and the delayed endpoint, appropriate representation of between-participant variation, and a suitable account of measurement and follow-up timing. Prediction for pending outcomes also relies on assumptions about why outcomes are pending. A scheduled assessment not yet due is different from an assessment missed because a participant became unwell or discontinued treatment.

Diagnostics, alternatives, and limitations

Before acting, examine whether the data being used resemble what the model can predict. Useful checks include observed-versus-predicted concentration and outcome trajectories, residual patterns over time, calibration of predictions at earlier data cuts when comparable observations have since matured, and sensitivity of the decision probability to plausible parameter or model changes. Investigate data-cut errors, unexpected visit timing, and whether the participants with mature outcomes differ systematically from those still pending.

A particularly important failure mode is treating delayed outcomes as ordinary missing observations without regard to their timing or cause. If participants with poor early response are more likely to discontinue or miss assessments, a model that predicts pending outcomes as though follow-up were unrelated to response may be optimistic. Conversely, predicting all pending participants as failures is also generally unjustified. The prediction model or the sensitivity analysis must address the actual follow-up process relevant to the decision.

Approach	Strength	Main limitation
Wait for outcomes to mature	Uses observed endpoints and avoids relying on predictions for pending outcomes.	Delays adaptation and may be impractical when outcome windows are long.
Use only outcomes already observed	Simple and transparent when the mature subset is representative and adequate.	Can be inefficient or biased if maturity is related to response, follow-up, or enrollment timing.
Posterior prediction from a longitudinal model	Uses available PK and partial trajectory information to quantify uncertainty about later outcomes.	Depends on model adequacy and assumptions about the outcome and follow-up processes.

Posterior prediction does not eliminate uncertainty or repair a poorly specified model. It can be sensitive to prior distributions, structural assumptions, sparse late-time observations, and extrapolation beyond observed exposure or follow-up. Decision thresholds also encode a trade-off: a rule that adapts quickly may be more sensitive to interim model error. The rule should therefore be evaluated under plausible trial scenarios, including delayed follow-up and model misspecification, before it is used.

Common mistake

Pitfall: “The participants with no recorded endpoint have not responded, so I will count them as failures,” or, conversely, “their concentrations look favorable, so I will count them as successes.” Both statements replace an unobserved outcome with a certainty unsupported by the data. The correction is to distinguish observed outcomes from pending outcomes, use a prespecified predictive model when prediction is warranted, and carry the resulting uncertainty into the adaptation rule.

Check Your Understanding

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

A participant has concentration measurements through week 2, but the primary outcome is assessed at week 8. At a week-4 interim, should the outcome be coded as zero, missing, or predicted? Explain.
A posterior distribution for a regimen's mean benefit has \(P(\Psi_A\geq\psi_{\min}\mid D_t)=0.74\). The prespecified continuation cutoff is 0.70. What does the rule say, and what does the probability mean?
State the two components of posterior predictive variance and explain why omitting one can affect an adaptation decision.
In the beta-binomial example, why is \(P(K\geq2)\) not the same quantity as \(P(p\geq0.5\mid D_t)\)?

Solutions. (1) The week-8 endpoint is pending at week 4; it is neither a recorded zero nor an observed failure. If the rule requires an interim decision before week 8, predict it using the specified model and retain predictive uncertainty. (2) The rule continues under the stated cutoff because \(0.74>0.70\). The value 0.74 is the posterior probability, conditional on the interim data and model, that the mean benefit meets the minimum; it is not a probability that the rule is correct. (3) Predictive variance is the expected conditional outcome variance plus the posterior variance of the conditional mean. Omitting the first ignores outcome variability; omitting the second ignores uncertainty in model quantities. Either omission can make predictions too certain and change whether a threshold is crossed. (4) \(P(K\geq2)\) is the predictive probability that at least two of four pending participants respond, integrating over uncertainty in \(p\). \(P(p\geq0.5\mid D_t)\) concerns the response probability itself, not the number of responses in a future group.