Why a Dose-Response Curve Needs More Than One Band
A simulated dose-response curve can answer different questions, and those questions do not have the same uncertainty band. One question is how uncertain the population mean outcome is at each dose, given uncertainty in model parameters. Another is how much outcomes may differ among individual patients at a dose. The first concerns uncertainty about a curve; the second concerns the distribution of outcomes around it. A band that does not specify which target it summarizes can be badly misinterpreted.
Here, dose \(d\) is a fixed treatment input, and \(Y(d;\theta,V,\epsilon)\) is a clinical outcome predicted by the QSP model. The vector \(\theta\) denotes uncertain model parameters, \(V\) denotes virtual-patient attributes drawn from the target population, and \(\epsilon\) denotes residual or measurement variation when that variation belongs in the clinical outcome being predicted. The endpoint, visit time, population, and treatment implementation are held fixed across doses. The target population mean at dose \(d\), conditional on a particular parameter vector, is
If the QSP output is a latent biological endpoint and the target is that latent endpoint, omit measurement error from \(\epsilon\). If the target is an observed clinical measurement, include the relevant residual variation. This choice changes the predictive distribution and must be made consistently at every dose.
Constructing a Pointwise Band for the Mean Curve
Suppose the model has \(M\) parameter draws \(\theta_1,\ldots,\theta_M\), with weights \(w_1,\ldots,w_M\) that sum to 1. The draws may come from a posterior distribution or from another justified uncertainty ensemble. At each dose, calculate the population mean conditional on each draw. If the target population is represented by \(N\) virtual patients, a direct estimate is
Use this expression only if the simulated outcomes include all variation that belongs in the conditional population mean. For a deterministic latent endpoint, the \(\epsilon_{ij}\) term is absent. If the endpoint has a known analytic conditional expectation, use that expectation rather than adding avoidable simulation noise. A weighted estimate of the ensemble-average mean is \(\sum_{j=1}^{M}w_j\widehat{m}(d,\theta_j)\). A pointwise uncertainty band is formed from the weighted lower and upper quantiles of the values \(\widehat{m}(d,\theta_j)\) at that dose.
The band is calculated separately at each dose, but the same parameter draws should be propagated across all doses. This preserves the relationship between a parameter draw’s predictions at different doses. The lower boundary at one dose and upper boundary at another do not necessarily belong to a single simulated parameter curve: the band is a collection of dose-specific summaries, not usually a simulated trajectory.
Worked Example: A Pointwise Band for the Mean Curve
Suppose ten equally weighted parameter draws have been simulated. For each draw, the QSP predictions have already been averaged over the same target virtual-patient population. The table gives the resulting population mean outcome at three doses, in hypothetical outcome units.
| Parameter draw | 0 mg | 5 mg | 10 mg |
|---|---|---|---|
| 1 | 0 | 5 | 10 |
| 2 | 0 | 6 | 12 |
| 3 | 1 | 6 | 13 |
| 4 | 1 | 7 | 15 |
| 5 | 2 | 8 | 17 |
| 6 | 2 | 9 | 19 |
| 7 | 3 | 10 | 21 |
| 8 | 3 | 11 | 23 |
| 9 | 4 | 12 | 26 |
| 10 | 5 | 14 | 30 |
At 10 mg, the ensemble-average population mean is
For illustration, use the linearly interpolated empirical quantile convention \(h=1+(M-1)p\), where \(p\) is the quantile probability and values are ordered from smallest to largest. At \(p=0.05\), \(h=1.45\), so the quantile is \(10+0.45(12-10)=10.9\). At \(p=0.95\), \(h=9.55\), so the quantile is \(26+0.55(30-26)=28.2\). Thus the pointwise 5th-to-95th percentile uncertainty band for the mean outcome at 10 mg is 10.9 to 28.2 outcome units. It summarizes the parameter ensemble’s uncertainty about the mean, not the range of outcomes expected among individual patients.
Why Mean-Curve Uncertainty and Patient Spread Differ
For an individual predictive outcome, both \(\theta\) and patient-level variables may vary. The law of total variance makes the distinction precise:
The first term is variation across parameter draws in their conditional mean outcomes. It is the variance component relevant to the mean-curve uncertainty described above. The second term is the average conditional outcome variance, which can include differences among patients and residual variation. A patient-level predictive distribution generally reflects both. A band for the mean curve should not simply be widened by adding the within-patient variance if its target is uncertainty in the mean.
This decomposition assumes a well-defined joint predictive model and finite variances. It does not imply that patient outcomes are normally distributed. In particular, QSP predictions may be skewed or multimodal, so predictive quantiles are often more informative than a normal approximation based on a mean and standard deviation.
Worked Example: Decomposing Mean Uncertainty and Patient Variation
At a hypothetical dose, suppose four equally weighted parameter draws imply conditional population means of 18, 20, 22, and 24 outcome units. Conditional on any one draw, suppose an individual patient’s outcome has variance \(6^2=36\) outcome-units squared. Treat these values as a simplified model for illustrating the variance decomposition.
The average of the four conditional means is \((18+20+22+24)/4=21\) outcome units. Their variance is
The average conditional patient-level variance is 36 outcome-units squared. Therefore the marginal variance for an individual outcome, when parameter uncertainty is also included, is \(5+36=41\) outcome-units squared, and its standard deviation is \(\sqrt{41}=6.403\) outcome units. The value 6.403 describes individual predictive spread under this simplified setup; it is not a half-width for the mean-curve band. Also, model parameters are shared across patients in a trial. Integrating over parameter uncertainty for an individual prediction does not mean that each patient in one simulated trial should be assigned an independent parameter draw.
Predictive Intervals at Individual Doses
To construct a predictive interval, simulate from the distribution matching the predictive question. For a new individual in the target population, this may involve drawing parameter uncertainty, drawing a virtual patient, and including residual variation if appropriate. At a fixed dose, sort the simulated outcomes and calculate the desired quantiles. State whether the interval concerns a new individual, a population summary, or another predictive unit; the same word “prediction” can otherwise conceal different targets.
Worked Example: An Individual Predictive Interval at 10 mg
Suppose 20 hypothetical simulated outcomes at 10 mg, generated to represent both parameter uncertainty and individual outcome variation, are sorted as follows. All values are in outcome units:
10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 24, 26, 28, 30, 33, 36, 40.
Use the same linear interpolation convention, now with \(M=20\). For the 5th percentile, \(h=1+19(0.05)=1.95\). Interpolating between the first two values gives \(10+0.95(11-10)=10.95\) outcome units. For the 95th percentile, \(h=1+19(0.95)=19.05\). Interpolating between the nineteenth and twentieth values gives \(36+0.05(40-36)=36.2\) outcome units. The empirical 90% predictive interval is therefore 10.95 to 36.2 outcome units. Under the simulation model, it summarizes the central 90% of individual outcomes at 10 mg; it is not a 90% interval for the population mean.
Diagnostics, Alternatives, and Limitations
A band is only as meaningful as the simulation target and the uncertainty ensemble behind it. Check that dose-specific simulations use the same endpoint definition, visit time, target population, and virtual-patient generation scheme. Inspect dose-response draws for implausible discontinuities, reversals, or numerical failures. A smooth-looking band does not by itself establish that the underlying QSP model is biologically or clinically adequate.
Monte Carlo error also needs attention. If the target mean is estimated by averaging \(N\) simulated patients for each parameter draw, a small \(N\) can make \(\widehat{m}(d,\theta_j)\) noisy. That noise can inflate the apparent spread across parameter draws and make a mean-curve band unstable. Increase simulation effort or use more precise conditional expectations, then check whether the estimated means and band quantiles stabilize. Monte Carlo error in estimating a band is not itself evidence of genuine parameter uncertainty.
For a quick summary, one might draw a curve from the ensemble-average means alone. This is useful for describing a central trend, but it hides uncertainty. A pointwise quantile band shows marginal uncertainty at each dose. It does not guarantee that the whole true curve lies between its boundaries with the stated probability: that would require a simultaneous band built from joint behavior across the dose range. Conversely, a patient-level predictive interval answers a different question and is often much wider than a mean-curve band.
The ensemble interpretation also matters. If parameter draws are weighted posterior draws, the band describes posterior uncertainty conditional on that model and analysis. If they are another plausible-parameter ensemble, its quantiles summarize that ensemble and its weights. Neither becomes a frequentist confidence band merely by being labelled a 90% or 95% band. Finally, uncertainty bands quantify uncertainty under the specified model and target; they do not account automatically for structural inadequacy, missing sources of variation, or unsupported assumptions about the simulated population.
A Practical Construction Sequence
Fix the endpoint, visit time, dose range, target population, and whether the target is a mean curve or an individual predictive distribution.
Use the justified parameter ensemble and retain the same parameter draws across doses. Draw virtual patients and residual terms only as required by the chosen target.
For a mean-curve band, calculate one conditional population mean per parameter draw and take quantiles across those means. For an individual predictive interval, take quantiles across simulated individual outcomes.
Assess Monte Carlo stability, report units and the quantile level, and state whether the summary is pointwise and which predictive target it represents.
Check Your Understanding
Answer each question before reading the brief solutions below it.
- At one dose, four equally weighted parameter draws give conditional population means of 8, 10, 12, and 14 units. What is their ensemble-average mean?
- In Question 1, the four conditional means have variance 5 units squared, and the average conditional individual-outcome variance is 16 units squared. What is the marginal variance and standard deviation for an individual outcome under the variance decomposition?
- A 90% pointwise band is constructed from parameter-specific population means. Does it imply a 90% probability that the whole dose-response curve lies inside the band? Why or why not?
- At 5 mg, a simulated 90% predictive interval for an individual is 4 to 20 units. Can it be reported as a 90% uncertainty interval for the population mean? Explain.
Solutions. (1) The average is \((8+10+12+14)/4=44/4=11\) units. (2) The marginal variance is \(5+16=21\) units squared, so the standard deviation is \(\sqrt{21}=4.583\) units. (3) No. The band is pointwise: it summarizes uncertainty separately at each dose and does not establish joint coverage of the entire curve. (4) No. It describes individual predictive outcomes under the stated simulation, including patient-level spread; it is not an interval for the population mean.