Why Propagate Uncertainty by Simulation?
A human prediction assembled from animal data depends on more than one uncertain quantity. An animal clearance estimate may be uncertain, the scaling exponent may be uncertain, and the human concentration model may be nonlinear in the resulting parameters. Translating only the animal point estimate and reporting one human concentration curve hides how those uncertainties combine.
Monte Carlo propagation makes this combination explicit. Each simulation draw represents one plausible set of animal estimates and scaling inputs. Those inputs are transformed into human parameters and passed through the prediction model. The resulting collection of predictions approximates the distribution implied by the specified input uncertainty and model.
The procedure does not make a translation model correct merely by running many draws. It propagates the uncertainty that was specified; it does not automatically include every source of uncertainty. In particular, do not interpret uncertainty across parameter draws as population variability among future patients unless that variability is explicitly represented.
From Animal Inputs to a Human Prediction
Let \(\boldsymbol{\theta}^{(m)}\) denote the inputs in simulation draw \(m\), including uncertain animal estimates and scaling quantities. A translation function \(g\) converts them to human parameters, and a prediction function \(f\) produces the quantity of interest at time \(t\):
For example, a clearance scaling relationship might be \(CL_H=CL_A(WT_H/WT_A)^b\), where \(CL_A\) is animal clearance, \(WT_H\) and \(WT_A\) are human and animal body weights, and \(b\) is an uncertain exponent. In each draw, sample the animal estimate and exponent from their specified distributions, calculate \(CL_H\), and use that clearance in the human PK model. Apply the same draw-by-draw logic to volume, potency, or other uncertain quantities.
The input distributions must represent the uncertainty relevant to the prediction. They may come from parameter estimates or posterior distributions established in earlier material on uncertainty in allometric scaling, parameter covariance, and Bayesian parameter uncertainty in human prediction. If inputs are dependent, the draws must respect the dependence represented by the chosen joint distribution; independently sampling dependent quantities changes the uncertainty being propagated. This tutorial focuses on carrying out and summarizing the propagation once an appropriate input distribution has been specified.
A practical implementation follows a short sequence:
Choose the prediction, such as concentration at a time point, a full concentration-time profile, or the probability that a response exceeds a threshold.
For each replicate, draw animal estimates and scaling parameters from the specified distributions, preserving any dependence they are assumed to have.
Transform that replicate's inputs to human parameters and calculate the target using the stated model. Keep the calculation internally consistent within each draw.
Use the simulated predictions to estimate quantities of interest, then assess simulation stability, implausible values, and whether the assumptions match the intended interpretation.
Worked Example: Simulating a Human Concentration
Worked Example: Simulating a Human Concentration
Suppose uncertain animal estimates and scaling inputs have been drawn and translated into the human volume \(V_H\) and clearance \(CL_H\). For this hypothetical illustration, the human prediction follows a one-compartment intravenous bolus model with dose \(D=100\) mg:
Assume \(t=2\) hours and five illustrative simulation draws yield the following translated human parameters. The different values represent different draws from the input uncertainty, not five observed people.
| Draw | \(V_H\) (L) | \(CL_H\) (L/h) | \(C(2)\) (mg/L) |
|---|---|---|---|
| 1 | 20 | 2.0 | 4.094 |
| 2 | 22 | 2.2 | 3.722 |
| 3 | 19 | 1.9 | 4.309 |
| 4 | 21 | 2.1 | 3.899 |
| 5 | 20 | 2.133 | 4.040 |
For draw 4, substitution gives:
which rounds to 3.899 mg/L. Applying the same equation to each draw gives unrounded predictions whose sum is approximately 20.0626 mg/L. Thus the simulated mean for these five draws is:
This mean is a Monte Carlo estimate of the mean prediction under the specified input distribution. Five draws are far too few for a stable description of most predictive distributions; they are used here only to show the arithmetic. In an actual analysis, use enough draws to make the reported summaries stable.
Summarizing the Predictive Distribution
The simulated predictions \(Y^{(1)},\ldots,Y^{(M)}\) can be summarized in several ways. The simulated mean estimates the expected prediction under the specified input distribution. The median and quantiles describe its location and spread without requiring a symmetric distribution. An interval formed from, for example, the 2.5th and 97.5th percentiles is a central 95% predictive interval under the specified simulation model: approximately 95% of that model's predictions fall between those quantiles. It is not automatically a confidence interval for a fixed parameter, nor does it include uncertainty sources omitted from the simulation.
For a decision threshold \(c\), the fraction of draws exceeding it estimates the probability of exceeding that threshold under the specified predictive distribution. The result is conditional on the input distributions and model, and should be reported with the target and units made clear.
The distinction is important. Increasing \(M\) generally reduces Monte Carlo error, but it does not narrow the underlying predictive distribution: that spread is determined by the input uncertainty and the model. A large simulation can estimate a wide predictive distribution precisely.
Worked Example: Reading Quantiles From Simulated Predictions
Suppose ten hypothetical simulations produce concentrations at a specified time, in mg/L. Sorted from lowest to highest, the values are 2.8, 3.1, 3.5, 3.7, 3.9, 4.2, 4.6, 5.0, 5.4, and 6.1.
With ten values, the median is the average of the fifth and sixth ordered values:
Using the nearest-rank convention, the empirical 10th percentile is the first ordered value because \(\lceil 10(0.10)\rceil=1\), and the empirical 90th percentile is the ninth ordered value because \(\lceil 10(0.90)\rceil=9\). The resulting empirical 10th-to-90th percentile range is 2.8 to 5.4 mg/L. It summarizes these ten draws only; with so few draws, it is a crude approximation to the corresponding quantiles of the full predictive distribution. A production analysis should use many more draws and check that its reported quantiles stabilize as \(M\) increases.
Worked Example: Probability of a Response Threshold
Worked Example: Probability of Exceeding a Response Threshold
Suppose a hypothetical pharmacodynamic model gives response \(R=100C/(EC_{50}+C)\), where \(C\) and \(EC_{50}\) have the same concentration units and response is expressed as a percentage of maximum. Six simulation draws produce the following \(C\) and \(EC_{50}\) values:
| Draw | \(C\) (mg/L) | \(EC_{50}\) (mg/L) | \(R\) (%) |
|---|---|---|---|
| 1 | 1.0 | 2.0 | 33.3 |
| 2 | 2.0 | 2.0 | 50.0 |
| 3 | 3.0 | 1.5 | 66.7 |
| 4 | 1.5 | 1.0 | 60.0 |
| 5 | 4.0 | 2.0 | 66.7 |
| 6 | 0.8 | 1.2 | 40.0 |
For draw 3, \(R=100(3.0)/(1.5+3.0)=66.7\%\). Four of the six simulated responses are at least 50%, so the estimated probability of reaching that response threshold is:
This is an empirical estimate under the six simulated draws, not a claim that the response in a particular future individual has a 66.7% chance of exceeding 50% unless the simulated distribution represents that individual's relevant uncertainty. Six draws also give a very imprecise estimate of a threshold probability; more simulations are needed for a reliable summary.
Justification and Practical Checks
The simulated mean is a natural estimator because, if each prediction draw has finite expectation, the average of independent draws has expectation equal to that prediction mean. If the prediction variance is finite, the variance of the average decreases in proportion to \(1/M\). Specifically, for prediction variance \(\sigma_Y^2\), independence gives:
The sample standard deviation \(s_Y\) estimates \(\sigma_Y\), giving the Monte Carlo standard error \(s_Y/\sqrt{M}\). The same principle explains why tail quantiles and small threshold probabilities usually require more draws than a mean: relatively few draws determine the tails or the rare event.
Check each simulation for failures that can be hidden by a plausible-looking average. Confirm that every draw uses the intended units, time scale, dose, and parameterization. Verify that translated quantities remain in their permitted domain, such as positive clearance and volume, and inspect unusually extreme predictions. If a model produces negative concentrations or numerical overflow, determine whether the input distribution permits invalid values or the calculation is unstable; do not silently discard such draws without a justified rule.
Also check Monte Carlo stability. Repeat the simulation with a larger \(M\), or compare summaries across independent runs. If the mean, quantiles, or threshold probability change materially, the simulation is not yet precise enough for the intended use. For a threshold probability estimated by a fraction \(\hat{q}\) from \(M\) independent draws, its approximate Monte Carlo standard error is \(\sqrt{\hat{q}(1-\hat{q})/M}\). This quantifies simulation error in that fraction, not uncertainty in the scientific model or in the input distributions.
Alternatives, Limitations, and a Common Pitfall
Monte Carlo propagation is flexible for nonlinear translation and prediction models, where an analytic distribution for the output may be unavailable. If a model is approximately linear and input distributions are well characterized, analytic propagation using variances and covariances may be faster and can make the dependence of output variance on inputs transparent. A first-order approximation can be inaccurate for strong nonlinearities or highly skewed inputs. Scenario analysis is another useful alternative: it evaluates a small number of deliberately chosen input combinations, such as low, typical, and high cases. Scenarios are interpretable but do not, by themselves, assign probabilities to outcomes or estimate a full predictive distribution.
Monte Carlo results are limited by the distributions supplied, the translation equations, and the prediction model. More draws reduce numerical simulation error but cannot repair biased animal estimates, unsupported scaling assumptions, or a model that is inappropriate for humans. A simulated interval therefore needs to be interpreted with its conditioning assumptions, rather than presented as a guarantee of clinical coverage.
- Pitfall: treating simulation size as evidence of model certainty. The wrong reasoning is, “I ran one million draws, so the human prediction is reliable.” A large \(M\) can make the computation precise while leaving the scientific assumptions weak.
- Correction: distinguish Monte Carlo error from uncertainty represented by the input distributions and from uncertainty not included in the model. Report the prediction summaries together with their assumptions and assess whether the simulation summaries are stable.
- Pitfall: translating averages instead of propagating draws. With nonlinear functions, prediction from average inputs generally differs from the average prediction: \(f(E[\boldsymbol{\theta}])\) need not equal \(E[f(\boldsymbol{\theta})]\).
- Correction: when the target is the predictive distribution, translate and predict each draw first, then summarize those predictions.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- Why can the prediction from mean input parameters differ from the mean of simulated predictions?
- A simulation yields a mean concentration of 8 mg/L and a predictive interval of 3 to 16 mg/L. Does increasing the number of draws necessarily make that interval narrower?
- For \(M=400\) independent simulations, 100 predictions exceed a specified threshold. Calculate the estimated exceedance probability and its approximate Monte Carlo standard error.
- Name two checks that can reveal a problematic simulation even when its average prediction seems plausible.
Solutions. (1) Nonlinearity means \(f(E[\boldsymbol{\theta}])\) need not equal \(E[f(\boldsymbol{\theta})]\); draw-by-draw prediction estimates the latter. (2) No. More draws improve estimation of the interval endpoints, but the underlying predictive spread is determined by the input distributions and model. (3) The estimated probability is \(100/400=0.25\). Its approximate Monte Carlo standard error is \(\sqrt{0.25(0.75)/400}\approx0.0217\), or 2.17 percentage points. (4) Examples include checking units and time scales, verifying that translated parameters are in valid ranges, inspecting extreme or impossible predictions, and comparing summaries across larger simulations or independent runs.