Choosing Regimens Across Population Variability

Why Compare Regimens Across Patients?

A regimen that produces a desirable exposure for the “average” patient may miss the target for many individuals. Clearance, body size, organ function, adherence, and other sources of variability spread predicted exposures across a population. The distribution matters: two regimens with the same mean exposure can have different proportions of patients below, within, or above a clinically relevant range.

A population comparison also requires a clear answer to whose outcomes count, and how much? A population-weighted average gives each patient the same weight, so groups contribute according to their prevalence. A decision-maker may instead assign explicit weights to groups or outcomes to reflect a stated clinical priority. Those weights should be visible in the analysis rather than hidden in a choice of summary statistic.

From an Exposure Distribution to Target Attainment

Suppose predicted exposure is lognormally distributed under a regimen: \(\ln C_r\sim N(\mu_r,\sigma_r^2)\). Here \(\mu_r\) and \(\sigma_r\) are the mean and standard deviation of log exposure, respectively. The median exposure is \(e^{\mu_r}\), and the geometric standard deviation is \(e^{\sigma_r}\). This is a modeling assumption, not a universal property of exposure data.

For a target range with \(0<L<U\), standardizing the log exposure gives the target-attainment probability. The calculation uses the standard normal cumulative distribution function \(\Phi\).

The two standardized limits locate the target boundaries in the distribution of log exposure. The difference between their cumulative probabilities is the fraction of the modeled population predicted to fall inside the target range. It is not the probability that a particular patient will attain the target after that patient’s exposure is known.

Population Weights and Decision Weights

When patients belong to distinct groups, estimate or predict attainment within each group before combining results. Let \(p_g\) be the proportion of the target population in group \(g\), and let \(a_{rg}\) be the predicted target-attainment probability for regimen \(r\) in that group. The population-average attainment is \(\sum_g p_g a_{rg}\), where the population proportions sum to one.

Population proportions answer a descriptive question: how would the regimen perform in the specified population mix? Decision weights answer a different question: how much should each group’s result count in a decision? If \(w_g\) are normalized decision weights, with \(\sum_g w_g=1\), a decision-weighted summary is \(\sum_g w_g a_{rg}\). It should not be described as the population-average attainment unless \(w_g=p_g\).

Weighting Patient Outcomes, Not Just Target Attainment

A target indicator assigns the same score to every patient within the target and the same score to every patient outside it. That can be too crude if outcomes differ in value or severity. Expected utility provides a general way to summarize predicted outcomes using explicit scores. For each patient or group, define a utility \(u(y)\) for outcome \(y\); the expected utility is the probability-weighted average of those scores. A population-level expected utility then averages across patients using the intended population or decision weights.

The scale and direction of utility must be stated. For example, a score of 1 might represent a substantial symptom improvement and 0 no improvement. Such scores are decision inputs, not facts produced by the exposure model. Their clinical rationale and sensitivity should be examined. This framework can summarize a single outcome dimension; it does not by itself resolve trade-offs among multiple objectives.

Justification, Assumptions, and Diagnostics

The weighted summaries follow from averaging patient-level outcomes. If patient \(i\) has an outcome score \(U_i\), the finite-population average is \(N^{-1}\sum_{i=1}^N U_i\). If patients are represented by groups with proportions \(p_g\), and the model gives the group-specific mean score \(E(U\mid g,r)\), then partitioning the population by group gives:

The same reasoning applies to target attainment by setting \(U_i=1\) when exposure is in the target and \(U_i=0\) otherwise. Replacing \(p_g\) with decision weights \(w_g\) is a separate normative choice, not a consequence of the population-average identity.

A defensible comparison depends on several conditions. The target population and its group proportions should be relevant to the decision. Exposure predictions should represent the intended regimen, dosing conditions, and sources of variability; the exposure-response relation and target range should be clinically justified. If a lognormal model is used, its fit should be checked rather than assumed. The comparison should also make clear whether uncertainty in model parameters and population proportions is represented or whether predictions are conditional on fixed estimates.

Useful diagnostics include comparing predicted exposure distributions with observed data, checking whether residuals or subgroup patterns suggest a poor distributional fit, and plotting the full exposure distributions alongside target boundaries. Compare predicted and observed target-attainment summaries where suitable data exist. Inspect subgroup-specific results: an overall mean can hide a poorly served group. Repeat the calculation under plausible alternative population proportions, target ranges, and utility scores. Large changes in the ranking signal that the decision depends on those choices and should be reported.

Alternatives, Limitations, and a Common Pitfall

Comparing mean or median exposure is simpler and can be useful for describing central tendency, but it does not show how much of the population lies within a target range. Target attainment is directly interpretable when the range is clinically justified, but it reduces a continuous exposure distribution to an inside-or-outside indicator. Expected utility retains distinctions among outcomes if meaningful scores are available, but its result depends on those scores. Reporting the distribution, subgroup results, and a weighted summary together is often more informative than relying on any one measure.

Model-based predictions are not direct observations for every patient. They can be sensitive to distributional assumptions, covariate representation, adherence assumptions, and uncertainty in estimated parameters. A high average attainment can coexist with low attainment in a subgroup, while a favorable decision-weighted score can reflect a deliberate priority rather than the population’s actual composition. These summaries support regimen comparison; they do not establish causal benefit beyond the evidence and assumptions underlying the models.

• Pitfall: choosing the regimen with the closest mean exposure to the target midpoint. That reasoning ignores distribution width and tails. A regimen centered at the midpoint may still have low attainment if between-patient variability is large. Calculate target attainment from the full predicted distribution.
• Pitfall: calling a decision-weighted result “population average.” If weights differ from population proportions, the summary answers a different question. Label the weights and state whose outcomes receive extra emphasis.
• Pitfall: treating utility scores as model outputs. The exposure or outcome model predicts probabilities; decision-makers supply the utility scale. Justify the scores and test whether the regimen ranking changes under plausible alternatives.