Why Build a Band Around a Concentration-Time Curve?
A simulated concentration-time curve is often presented as though it were known precisely. In practice, however, the model parameters are estimated, and plausible parameter values can produce meaningfully different curves. An uncertainty band displays that variation over time. It can show, for example, whether uncertainty is concentrated near the peak, grows late in the dosing interval, or changes the apparent timing of a concentration threshold.
A band is not automatically a prediction range for individual patients. The target here is uncertainty in a population concentration-time summary, such as the population mean concentration under a specified dosing schedule and target covariate distribution. The construction must preserve that target: for each parameter draw, first obtain the intended population summary curve, then aggregate those curves across draws.
The curve depends on the estimand as well as the model. A mean for a specified covariate profile is not the same target as a mean averaged over a population covariate distribution. Likewise, a mean concentration curve is different from a population percentile curve. State the dosing schedule, covariate target, population summary, and time range before constructing the band.
Constructing Pointwise Bands
Suppose \(B\) parameter draws are available, denoted \(\theta^{(1)},\ldots,\theta^{(B)}\). They may come from a nonparametric bootstrap or from a Bayesian posterior, as in the earlier material on parameter uncertainty from bootstrap and posterior draws. For each draw, simulate or calculate the chosen population summary over the same time grid \(t_1,\ldots,t_J\). Write the resulting curve as \(m_b(t_j)=m(t_j;\theta^{(b)})\).
At a particular time \(t_j\), the \(B\) values \(m_1(t_j),\ldots,m_B(t_j)\) form an empirical distribution of the summary across parameter draws. Its chosen lower and upper quantiles provide a pointwise interval at that time. Repeating this calculation at every grid time produces a pointwise band.
Here \(Q_p\) is the empirical \(p\)-quantile under a stated quantile convention. For bootstrap draws, these limits are commonly used as a pointwise bootstrap confidence interval for the population summary at time \(t_j\). For posterior draws, they are pointwise posterior credible limits. These are not interchangeable interpretations: the bootstrap interval has a repeated-sampling confidence interpretation under appropriate bootstrap validity, while a posterior interval summarizes posterior probability conditional on the model and prior.
Worked Example: A Pointwise Band at One Time
Suppose ten hypothetical bootstrap parameter draws produce population mean concentrations at 4 hours, in mg/L, with sorted values \(3.8, 4.1, 4.3, 4.5, 4.6, 4.8, 5.0, 5.2, 5.5, 5.9\). Use the nearest-rank empirical quantile convention, \(Q_p=x_{(\lceil Bp\rceil)}\), to construct a pointwise 80% interval.
For an 80% interval, the tail probability is \(\alpha/2=0.10\). The lower rank is \(\lceil 10(0.10)\rceil=1\), so \(L(4)=3.8\) mg/L. The upper rank is \(\lceil 10(0.90)\rceil=9\), so \(U(4)=5.5\) mg/L.
Thus, the pointwise 80% bootstrap interval at 4 hours is 3.8 to 5.5 mg/L. It concerns the population mean concentration at that time, not the range expected to contain 80% of individual subjects. With only ten draws, the empirical limits are coarse; real analyses generally require substantially more draws for stable tail quantiles.
A pointwise band does not promise that the entire true curve lies between its limits with the stated confidence level. Each time is handled separately. Even if every pointwise interval has 95% coverage under its intended interpretation, the probability that all intervals cover simultaneously is generally less than 95%.
Simultaneous Bands Across Time
When the question concerns the curve as a whole—for example, whether a candidate curve remains within an acceptable range throughout a dosing interval—a simultaneous band is more appropriate. One common construction uses a maximum standardized deviation. Let \(\widehat m(t_j)\) be the reference curve, such as the curve at the fitted parameter estimate for a bootstrap analysis, and let \(s(t_j)>0\) be a scale for deviations at that time. For each draw, calculate the largest standardized absolute deviation across the grid.
Let \(c\) be the selected upper quantile of the empirical distribution of \(M_1,\ldots,M_B\). A simultaneous band on the grid is then \(\widehat m(t_j)\pm c\,s(t_j)\). The scale may be the bootstrap standard deviation of the curve at each time, or another justified positive scale. Because the maximum is computed across all grid points for each draw, this approach accounts for the dependence among times represented in the simulated curves.
Worked Example: A Simultaneous Band on a Three-Time Grid
Consider a hypothetical fitted population mean curve \(\widehat m(t)\) with concentrations \((10,6,3)\) mg/L at 1, 4, and 8 hours. Suppose the scales \(s(t)\) at those times are \((1,0.5,0.25)\) mg/L. Five hypothetical bootstrap curves have deviations from the fitted curve, in the same time order:
Draw 1: \((-1,0.2,0.1)\); draw 2: \((0.5,-0.4,0.25)\); draw 3: \((1,0.1,-0.1)\); draw 4: \((-0.5,0.3,-0.2)\); draw 5: \((0.2,-0.1,0.05)\), all in mg/L.
For draw 1, the standardized absolute deviations are \((1,0.4,0.4)\), so \(M_1=1\). For draw 2 they are \((0.5,0.8,1)\), so \(M_2=1\). The corresponding calculations give \(M_3=1\), \(M_4=0.8\), and \(M_5=0.2\). Using the largest observed maximum as an illustrative critical value gives \(c=1\).
The resulting grid band is the fitted curve plus or minus one scale unit: at 1 hour, \(10\pm1=(9,11)\) mg/L; at 4 hours, \(6\pm0.5=(5.5,6.5)\) mg/L; and at 8 hours, \(3\pm0.25=(2.75,3.25)\) mg/L. With only five draws, this is a demonstration of the calculation, not a reliable calibrated 95% band. In an analysis, the quantile level should match the desired simultaneous coverage and be estimated from enough draws.
For posterior draws, a related construction can center deviations on a posterior summary curve and take a posterior quantile of the maximum standardized deviation. Its interpretation is a joint posterior credible band, conditional on the model and prior, not a frequentist confidence band. In either setting, the band is simultaneous only over the grid used to calculate the maxima. A coarse grid can miss a sharp peak or a crossing between grid times, so refine the grid where the curve changes rapidly and check whether the band changes materially.
Keep Parameter Uncertainty Separate From Subject Variation
A frequent source of error is pooling every simulated subject concentration across every parameter draw and treating the resulting quantiles as an uncertainty band for a population mean curve. That pooled distribution includes between-subject variation as well as differences induced by parameter draws. It answers a different question, closer to the distribution of individual outcomes under the simulation design.
Worked Example: Average Within Each Draw Before Aggregating Curves
At 2 hours, suppose three hypothetical parameter draws each generate concentrations for three new subjects. Draw 1 gives \(8,10,12\) mg/L; draw 2 gives \(9,11,13\) mg/L; draw 3 gives \(7,8,9\) mg/L.
For the population mean summary, first average within each parameter draw. The three curve values are \((8+10+12)/3=10\) mg/L, \((9+11+13)/3=11\) mg/L, and \((7+8+9)/3=8\) mg/L. The uncertainty distribution for the mean curve at this time is therefore represented by \(8,10,11\) mg/L across parameter draws.
By contrast, pooling all nine subject values gives \(7,8,8,9,9,10,11,12,13\) mg/L, which reflects both subject-to-subject differences and parameter-draw differences. Its spread is not the uncertainty in the population mean. In an actual simulation, more subjects per parameter draw may be needed to estimate each population mean accurately; the within-draw subject simulations are used to calculate the summary, not treated as extra parameter draws.
The distinction also determines what should vary across draws. If the target is parameter uncertainty in a population mean, draw the model parameters jointly and calculate that mean under the specified random-effect and covariate distributions. If the target is a population percentile, calculate that percentile within each parameter draw and then summarize its variation across draws. The target summary must be calculated before taking uncertainty quantiles.
Assumptions, Diagnostics, and Alternatives
A simulation-based band is only as credible as the draws and curve calculation behind it. The parameter draws should represent the intended uncertainty distribution and preserve dependence among parameter estimates. The model, dose schedule, time units, covariate distribution, random-effect distribution, and summary definition should be consistent across draws. Curves should be evaluated at common times, and any numerical failures or implausible parameter values should be investigated rather than silently discarded.
Useful checks include plotting a sample of individual draw-curves alongside the central curve, checking that concentration units and time points agree, and inspecting the width of the band over time. Sudden spikes in width may reflect genuine parameter sensitivity, but can also signal unstable simulations, extreme draws, or a summary that is poorly estimated. If random subjects are simulated to approximate the population mean, repeat or enlarge that inner simulation to check that the estimated mean curve is not dominated by subject-simulation noise.
Choose the construction according to the scientific question:
| Approach | What it summarizes | Key limitation |
|---|---|---|
| Pointwise quantile band | Uncertainty in the population summary separately at each time | Does not provide joint coverage for the entire curve |
| Maximum-deviation band | Uncertainty in the curve jointly over a specified time grid | Can be wider, and its coverage applies to the chosen grid and calibration |
| Prediction interval across subject outcomes | Variation in individual future outcomes under the simulation design | Includes subject variability and is not a parameter-uncertainty band for a population summary |
No band removes model limitations. A narrow band can still be misleading if the model structure or target population is inappropriate, or if relevant uncertainty was not represented in the parameter draws. A band also does not establish that every concentration curve between its limits is equally plausible; it summarizes uncertainty according to a particular construction.
Common Mistakes
- Calling pointwise limits a whole-curve guarantee. Quantiles computed separately at each time do not ensure that the entire curve is covered at the stated level. Use a simultaneous construction when the question is joint over time.
- Pooling subjects and parameter draws. This mixes individual variability with uncertainty in the population summary. Calculate the target summary within each parameter draw first.
- Changing the target across draws. Using different covariate profiles or summary definitions produces curves that are not comparable as draws of one estimand.
- Interpreting bootstrap and posterior bands identically. Bootstrap confidence and posterior credible intervals have different inferential meanings, even if their endpoints look similar.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A pointwise 90% interval is calculated at every time on a concentration-time grid. Does this imply 90% coverage of the entire curve? Why or why not?
- At one time, four parameter draws give population mean concentrations of 4, 5, 7, and 8 mg/L. Under the nearest-rank convention, what is the 50% interval using the 25th and 75th percentiles?
- Why should simulated subject concentrations usually be summarized within each parameter draw before constructing a band for the population mean?
- In a maximum-deviation construction, what does the maximum over time accomplish?
Solutions. (1) No. Each interval is calibrated separately at a time; joint coverage across all times is generally lower and requires a simultaneous method. (2) With \(B=4\), the nearest-rank indices are \(\lceil4(0.25)\rceil=1\) and \(\lceil4(0.75)\rceil=3\), so the interval is 4 to 7 mg/L. (3) Averaging within each draw estimates the population mean for that parameter draw; pooling subjects would also include between-subject variability. (4) It records each draw's largest standardized departure anywhere on the grid, allowing the critical value to calibrate a band for the curve jointly rather than separately at each time.