Deriving Posterior Target-Attainment Rules

From an Interim Model to an Adaptation Rule

An interim pharmacometric model can describe uncertainty about exposure or response, but an adaptation requires a decision statement: what quantity must attain what target, and how much posterior support is required before acting? A rule based only on a fitted mean can conceal uncertainty. A posterior target-attainment rule makes that uncertainty explicit by calculating the probability of a prespecified target event given the interim data.

The target event must be defined before applying the rule. It may concern a model parameter, such as whether a population mean response exceeds a minimum, or a future observation, such as whether a patient's exposure lies within a therapeutic interval. These are different probabilities. The first concerns uncertainty about a population quantity; the second also includes the variability of future observations around that quantity.

For a parameter target, the probability is calculated from the posterior distribution of that parameter. For a future-patient target, it is a posterior predictive probability: uncertainty about model parameters is combined with the conditional variability of a future observation. The distinction matters because a posterior mean exposure inside a target interval does not imply that most future patients will fall inside that interval.

Deriving a Predictive Exposure Probability

Suppose log exposure \(X\), under a fixed adaptation, follows \(X\mid\theta\sim N(\theta,\sigma^2)\), where \(\theta\) is the population mean log exposure and \(\sigma^2\) is residual variance among future patients. Suppose the interim data give \(\theta\mid D\sim N(m,v)\). These assumptions describe a normal population on the log scale, a known residual variance, and a normal posterior for the mean. Conditional on \(\theta\), the future exposure is independent of the interim data.

The probability of attaining an exposure interval \([L,U]\) is obtained by integrating over uncertainty in \(\theta\). Because \(X=\theta+\varepsilon\), with \(\varepsilon\sim N(0,\sigma^2)\) independent of \(\theta\mid D\), the posterior predictive distribution adds the two variances:

Here \(\Phi\) is the standard normal cumulative distribution function. The equation follows because \(L\leq e^X\leq U\) is equivalent to \(\log L\leq X\leq\log U\), and the probability of a normal variable lying between two limits is the difference of its cumulative probabilities at those limits. The term \(v\) represents uncertainty about the mean exposure; \(\sigma^2\) represents future-patient variability. Omitting either term changes the target event being evaluated.

Parameter Targets for Response

A response target may instead be defined on a population parameter. For example, the event may be that the population mean response \(\theta\) exceeds a clinically relevant minimum \(\theta_0\). A simple derivation is available when normally distributed responses have known variance and the prior for \(\theta\) is normal. If \(Y_i\mid\theta\sim N(\theta,\sigma^2)\) independently and \(\theta\sim N(m_0,\tau_0^2)\), then after observing \(n\) responses with sample mean \(\bar{x}\), the posterior variance and mean are:

These expressions result by multiplying the normal prior density by the normal likelihood and collecting the terms in \(\theta^2\) and \(\theta\) in the exponent. The resulting posterior is normal with precision \(1/v_n=1/\tau_0^2+n/\sigma^2\). For the one-sided target \(A=\{\theta\geq\theta_0\}\), the attainment probability is \(1-\Phi((\theta_0-m_n)/\sqrt{v_n})\). This is a probability about the population mean, not a predictive probability for an individual patient's response.

What a Threshold Does—and Does Not—Mean

A threshold rule is a decision policy, not a guarantee of success. The threshold \(\gamma\) expresses how much posterior support the policy requires for a defined event before an adaptation qualifies. Its value should be chosen in advance with the clinical consequences of acting and not acting in view. It should not be described as a frequentist type I error rate: a rule requiring a posterior probability of 0.90 does not, by itself, ensure a 10% false-positive rate across repeated trials.

One way to justify a threshold is through expected utility. Suppose taking an adaptation yields incremental utility \(B>0\) if the target is attained and incurs incremental utility \(-C\), with \(C>0\), if it is not attained. Relative to not taking that adaptation, the expected utility is \(qB-(1-q)C\). Acting is preferred when that quantity is positive, which gives the probability threshold below.

Applying the Rule and Checking Its Weak Points

A reproducible decision rule links the event, posterior calculation, threshold, and operational consequence. Before the interim analysis, specify the target and its units, the population or future observation to which it refers, the model used to calculate its probability, the threshold, and what happens if no action qualifies. If the rule includes several candidate actions, define in advance how to proceed when several meet the threshold, rather than treating the threshold alone as a complete ranking rule.

Useful diagnostics include checking that the target event matches the clinical question, verifying transformations and units, and confirming that numerical integration or simulation estimates the intended posterior probability with adequate precision. Posterior predictive checks can reveal whether the assumed exposure or response distribution reproduces important features of observed data. Sensitivity analyses should examine plausible alternative priors, residual variability, model parameters, and thresholds. If a small reasonable change moves the probability across \(\gamma\), the decision is fragile even if the calculation is correct.

A posterior mean compared with a target is a simpler alternative, but it ignores uncertainty: two analyses with the same mean and different posterior spreads can produce different target probabilities. A credible interval can summarize parameter uncertainty, but a rule such as “the interval lies above the target” is not identical to a specified posterior probability threshold unless its relationship to that threshold is established under the model. For future-patient exposure, neither a parameter mean nor a parameter interval substitutes for the predictive distribution. A fixed, non-adaptive rule avoids interim posterior calculations but cannot respond to interim evidence in the same way.

The calculation is only as reliable as its target definition and model. Unmodeled heterogeneity, misspecified tails, inappropriate borrowing of information, or an interim sample unrepresentative of the future population can distort attainment probabilities. Further, evaluating many candidate actions and acting on the most favorable result is not automatically accounted for by the probability threshold. The entire adaptive policy should be assessed under plausible trial scenarios, including model departures, to understand its operating characteristics. That assessment is separate from the meaning of any single posterior probability.

Common Pitfall