Propagating Parameter Uncertainty to Clinical Decisions

From Parameter Uncertainty to a Clinical Decision

An exposure-response model can predict different response rates at the same dose because its fitted parameters are uncertain. A clinical decision, however, requires comparing doses using more than a single best-fit prediction: it must account for how that uncertainty changes predicted benefit and the balance of benefit against dose-related burden. Propagating parameter uncertainty means evaluating the model across draws representing uncertainty in its parameters, then summarizing the resulting predictions and decisions.

This is not the same as saying that patients have uncertain outcomes because of parameter uncertainty. Even if a response probability were known exactly, individual patients could still respond or not respond. Here the focus is uncertainty about the model’s response probability and how it affects a decision. The earlier material on distinguishing variability from decision uncertainty is useful for keeping those two sources conceptually separate.

The draws must have a clear interpretation. They might, for example, be posterior draws under a specified Bayesian model or bootstrap estimates reflecting repeated-sampling uncertainty under a specified resampling procedure. These are not interchangeable interpretations, and neither turns a model-based prediction into an observed clinical outcome. In the worked case below, the draws are treated as equally weighted representations of parameter uncertainty for teaching purposes. They are deliberately few, so they illustrate arithmetic and decision summaries rather than provide a reliable real-world uncertainty analysis.

Model, Assumptions, and Decision Rule

Suppose the modeled outcome is a binary clinical response and dose \(d\) is measured in milligrams. For this example, use a logistic exposure-response model with a fixed intercept and an uncertain slope:

The slope \(\beta_1\) is measured per milligram, so \(\beta_1d\) and the linear predictor are dimensionless. A positive slope means the modeled response probability rises with dose. This simple setup isolates uncertainty in one parameter; it is not a general representation of all fitted exposure-response models. The calculations assume that the model form is adequate for the decision being considered, the candidate doses are clinically feasible, and the parameter draws represent the uncertainty to be propagated. They do not account for uncertainty in the model structure, unmodeled discrepancies, or variation in patient characteristics.

To make the choice explicit, suppose utility is measured in hypothetical decision points, with 100 points assigned to a response and a dose burden of \(0.5d^2\) points. If the modeled response probability is \(p\), the expected utility at dose \(d\) is

This is a deliberately simple preference rule, not a clinical recommendation. The dose term represents a stipulated burden; its coefficient and functional form would need justification for an actual decision. Given parameter draws, the mean utility for dose \(d\) is the average of its draw-specific utilities. Choosing the dose with the highest mean utility is appropriate only if this is the agreed decision criterion and the draws represent the uncertainty relevant to that choice.

Summarizing Predictions Across Draws

For equally weighted draws, the mean predicted response probability at a dose is

This is an average over parameter uncertainty. Because the logistic function is nonlinear, it generally differs from evaluating the model at the average parameter value. In this example, the slope draws are \(0.15,\ 0.20,\ 0.27,\ 0.32,\) and \(0.40\) per mg. Each draw is evaluated at every candidate dose, and then the predictions are averaged. The response model is increasing in dose for every listed draw because all slopes are positive.

From Predictions to Dose Comparisons

For each draw, calculate utility at every candidate dose using the same decision rule. The mean utility is the average of these draw-specific utilities. Since the dose burden in this example does not depend on the uncertain slope, mean utility can also be computed as \(100\bar p(d)-0.5d^2\). That equality follows by averaging the utility formula:

A useful additional summary is the probability, across the uncertainty draws, that a dose has the highest utility among those considered. With equally weighted draws, this is the proportion of draws for which that dose is best. It is conditional on the model, candidate options, utility rule, and interpretation of the draws. It is not the probability that the dose will be clinically successful, nor does it include unrepresented uncertainties.

Justification, Diagnostics, and Alternatives

The propagation calculation is a direct application of evaluating a function over parameter uncertainty: for each draw, insert its parameter value into the model, obtain the prediction, and then summarize the resulting values. This order matters. For nonlinear models, averaging parameters first and calculating one prediction can discard meaningful variation in predictions. In general, if \(g\) is nonlinear, the average of \(g(\theta^{(m)})\) need not equal \(g\) evaluated at the average of the draws. The appropriate summary depends on the question: a prediction at a central parameter estimate answers a different question from the average prediction across parameter uncertainty.

Before interpreting decision summaries, check the calculation and its scope. Confirm that doses use the model’s stated units; verify that each draw yields a valid probability between 0 and 1; and check the utility arithmetic and candidate-dose comparison. Inspect predictions over the clinically relevant dose range for implausible trends or extrapolation beyond the data that informed the model. A sharp change in dose ranking across modestly different draws is a sign that the decision is sensitive to parameter uncertainty, not proof that the model is wrong.

The numerical summaries are also sensitive to how the parameter uncertainty is represented. A small number of draws can give unstable means and especially coarse proportions of draws in which a dose is optimal. In practice, use enough well-generated draws for stable summaries and examine their precision. If draws are weighted unequally, use the specified weights rather than treating every draw as equally likely. Do not create plausible-looking draws by varying parameters arbitrarily and then interpret their frequencies as probabilities.

A point-estimate approach is a simpler alternative: predict at one fitted parameter estimate, then compare dose utilities using those predictions. It is easy to communicate, but hides how parameter uncertainty might change predictions or dose rankings. Interval summaries for response rates can communicate a range, but by themselves do not say which dose is preferred; the utility rule connects predictions to the decision. A deterministic scenario analysis can show consequences under selected parameter values, but scenario frequencies are not probabilities unless the scenarios and weights have a justified uncertainty interpretation.

This propagation exercise is conditional on the model and uncertainty representation provided. It does not establish that the exposure-response relationship is causal, that the model is correct, or that a dose is safe or effective. It also does not automatically include patient-level variability, uncertainty in the utility preferences, or uncertainty from alternative model structures. Decision conclusions should be reported with the model, dose options, uncertainty source, and utility assumptions that produced them.

Common Pitfall

• Wrong reasoning: “The model predicts a 34.72% response rate at 5 mg, so about 35% of future patients are certain to respond, and 5 mg is the clinically optimal dose.”
• Correction: The 34.72% is an average model-based probability across the specified parameter draws, not a certainty about a future sample. The 5 mg choice is optimal only among the listed doses under the stipulated utility rule and uncertainty representation. It does not establish clinical optimality outside those assumptions.
• Another error: “The dose that is best in 60% of draws has a 60% chance of success.” The correct interpretation is that it had the highest calculated utility in 60% of these equally weighted parameter draws. That is not the probability of a patient response or trial success.