From a Calculated Dose to a Reliable Recommendation
A dose equation can return many decimal places even when its inputs are uncertain. That numerical precision describes the arithmetic, not how reliably the dose will achieve its intended exposure. In special populations, uncertainty in clearance, body size, maturation, or kidney function can materially change the dose implied by an exposure target.
The distinction matters in practice. If a model gives a dose of 35.37 mg per interval, that is a precise result for the values entered. It is not, by itself, evidence that 35.37 mg is the uniquely correct dose for a patient. A sound interpretation asks how uncertain the inputs are, how those uncertainties propagate through the dose equation, and whether the resulting range is narrow enough to support the intended decision.
The calculations below use a target average steady-state concentration \(C_{\mathrm{target}}\), dosing interval \(\tau\), bioavailability \(F\), and clearance \(CL\). For a linear model, the dose per interval that corresponds to the target is \(D=C_{\mathrm{target}}CL\tau/F\). The equation gives a plug-in dose when \(CL\) and the other inputs are fixed at their estimated values; uncertainty analysis treats uncertain quantities as uncertain rather than exact.
Propagating Uncertainty Through Clearance
Suppose clearance is represented by a body-size, maturation, and reference-clearance model:
Here \(W\) is weight in kilograms, \(CL_{\mathrm{ref}}\) is reference clearance in L/h, \(\theta\) is the allometric exponent, and \(M\) is a dimensionless maturation factor. This equation is useful for illustrating how uncertainty in the components enters the dose calculation. It does not establish that the model is appropriate for every drug or patient group.
A common local approximation is to work on the log scale. Taking logarithms gives \(\ln(CL)=\ln(CL_{\mathrm{ref}})+\theta\ln(W/70)+\ln(M)\). For small uncertainties, the variance of a function can be approximated by its local derivatives and the covariance matrix of its inputs. If the log-scale inputs are independent, their contributions add. If inputs are correlated, covariance terms must also be included; ignoring them can either overstate or understate uncertainty.
For the clearance model, if \(CL_{\mathrm{ref}}, W, \theta,\) and \(M\) are independent, the corresponding first-order log-clearance variance is approximately \(\operatorname{Var}(\ln CL_{\mathrm{ref}})+\theta^2\operatorname{Var}(\ln W)+[\ln(W/70)]^2\operatorname{Var}(\theta)+\operatorname{Var}(\ln M)\). Because \(D=C_{\mathrm{target}}CL\tau/F\), uncertainty in \(\ln(D)\) is the same as uncertainty in \(\ln(CL)\) when target, interval, and bioavailability are fixed. If any of those quantities are also uncertain, their contributions and any relevant covariances must be included.
Worked Example: Clearance and Dose With Several Uncertain Inputs
Worked Example: Clearance and Dose With Several Uncertain Inputs
Hypothetical calculation. Consider a child weighing 5.0 kg, with estimated maturation factor \(M=0.40\). Use \(CL_{\mathrm{ref}}=6.0\) L/h, \(\theta=0.75\), \(C_{\mathrm{target}}=8.0\) mg/L, \(\tau=12\) h, and \(F=0.90\). The nominal model inputs give:
The plug-in dose per interval is therefore
The units cancel to milligrams, as required. This result is the dose calculated from the specified inputs; it does not yet express uncertainty.
Now suppose, hypothetically, that the independent uncertainties are summarized by log-scale variances of 0.009950 for \(CL_{\mathrm{ref}}\), 0.000225 for weight, and 0.03922 for \(M\). Suppose the standard error of the allometric exponent is 0.08. The exponent’s contribution to log-clearance variance is
With independence, the approximate total is
If the log-dose uncertainty is approximately normal, an illustrative 95% uncertainty interval is the nominal dose multiplied by \(\exp(\pm1.96SE_{\ln D})\):
This interval is wide compared with the point estimate. Under the stated uncertainty distributions and assumptions, it summarizes plausible dose values implied by the uncertain inputs; it does not mean there is a 95% probability that a fixed, unknown parameter lies in this particular interval. If the inputs’ uncertainty summaries are frequentist standard errors, the interpretation depends on the validity of the approximation and how those estimates were obtained. If they describe probability distributions for the inputs, the interval is conditional on those distributions and the model.
The calculation relies on more than arithmetic. It assumes the clearance model is suitable at the child’s age and size, the uncertainty summaries are meaningful for the inputs, the inputs are independent, and the log-dose distribution is adequately approximated by a normal distribution. The numerical interval should not be treated as reliable merely because a formula produced it.
Kidney-Function Uncertainty and a Dose Interval
A kidney-function estimate may enter a dose calculation through a clearance model. Its uncertainty can be propagated just like uncertainty in weight or maturation, but the relationship between the estimate and drug clearance must be specified. The next example uses a deliberately simple linear relationship to make the arithmetic transparent; the coefficient is hypothetical and is not a general renal-dose rule.
Worked Example: Uncertainty From an Estimated Kidney Function
Hypothetical calculation. Suppose a linear model for a drug gives \(CL=1.2+0.05G\) L/h, where \(G\) is kidney function in mL/min. Assume a patient’s estimated \(G\) is 30 mL/min with standard error 6 mL/min. Treat the model coefficients as fixed for this example, and suppose the estimate’s sampling uncertainty is approximately normal.
The plug-in clearance is
Because the model is linear, the standard error of estimated clearance is exactly \(0.05(6)=0.30\) L/h under these assumptions. With \(C_{\mathrm{target}}=8.0\) mg/L, \(\tau=12\) h, and \(F=1\), the plug-in dose and its propagated standard error are
A normal-approximation 95% interval is
This interval quantifies uncertainty propagated from the kidney-function estimate alone. It does not include uncertainty in the model coefficient, clearance unrelated to kidney function, the target, or other factors omitted from the example. The normal approximation also assigns some probability to values below zero for kidney function; here that probability is negligible at 30 with standard error 6, but a normal approximation would need reconsideration if the estimate were near zero.
An estimate’s standard error is not the same as the patient’s biological variation over time, nor is it automatically a complete measure of error in the underlying kidney function. It characterizes sampling uncertainty under its estimation method. Bias, equation choice, measurement conditions, and model misspecification can add uncertainty that the reported standard error does not capture.
Why Dependence Between Inputs Matters
The independence assumption in a simple variance sum is consequential. Weight and maturation estimates may be related, and estimates derived from the same data can have correlated errors. On the log scale, for uncertain \(\ln(W)\) and \(\ln(M)\), the variance includes the term \(2\theta\operatorname{Cov}(\ln W,\ln M)\). Positive covariance increases the variance when both inputs raise clearance; negative covariance decreases it. A covariance should be based on the way the estimates were obtained, not chosen simply to narrow an interval.
Worked Example: Correlated Weight and Maturation Uncertainty
Hypothetical calculation. Return to the nominal dose of 35.3708 mg per interval. Suppose the log-scale standard deviations for weight and maturation are 0.05 and 0.20, respectively, and their log-scale correlation is \(-0.50\). Treat the other inputs as fixed, and assume this joint uncertainty is approximately normal on the log scale.
The variance of log dose is the variance of log clearance. The weight contribution is \(0.75^2(0.05)^2=0.00140625\), and the maturation contribution is \((0.20)^2=0.04000\). The covariance term is
Thus,
The corresponding approximate 95% interval is
The negative covariance narrows the interval relative to the same marginal uncertainties treated as independent. If the correlation were positive instead, the covariance term would increase the variance. This is not evidence that negative dependence should be assumed; it shows why a defensible joint uncertainty model matters.
Checking Whether the Uncertainty Analysis Is Informative
Before using a propagated interval to support a recommendation, check the inputs and the approximation:
- Verify the uncertainty summaries. Establish whether each value is a standard error, a confidence-limit-derived quantity, a coefficient of variation, or another uncertainty description. Do not combine these as if they were interchangeable.
- Check the model domain. Confirm that the clearance relationship is supported for the relevant age, size, maturation range, and kidney function. An interval cannot repair a model applied outside its useful domain.
- Consider dependence. Identify whether inputs share data, equations, or measurement processes. If they do, use their covariance when available or state clearly that independence is an assumption.
- Assess the approximation. Strong nonlinearity, large uncertainty, or a variable close to a boundary can make a first-order or normal approximation misleading. A simulation using defensible joint input distributions can be more appropriate in such cases.
- Compare the range with the decision. Ask whether the uncertainty range would imply materially different clinical choices. If it spans doses with meaningfully different exposure consequences, the point estimate alone is not a sufficient basis for confidence.
An interval represents only the sources of uncertainty included in its construction. For example, an interval based on estimated weight and maturation does not automatically account for model misspecification, unrepresented developmental changes, or uncertainty in the exposure target. Those omissions should be stated rather than hidden behind extra decimal places.
| Approach | What it provides | Main caution |
|---|---|---|
| Plug-in calculation | A dose calculated from selected point estimates | Does not describe input uncertainty |
| Local uncertainty propagation | A compact approximation based on derivatives and a covariance matrix | Can be inaccurate with strong nonlinearity or large uncertainty |
| Simulation from input distributions | A distribution or interval for dose under specified joint input distributions | Results depend on the distributions and model supplied |
| Scenario calculations | Shows dose results under selected alternative assumptions or input values | Scenarios are not a probability interval unless probabilities are assigned |
Common Pitfall: Reporting Precision as Reliability
Wrong reasoning: “The dose calculation is 35.37 mg, so the dose is known accurately to within 0.01 mg.” The displayed precision only reflects the arithmetic performed on the chosen inputs. In the first example, plausible input uncertainty produced an approximate interval from 19.4 to 64.5 mg per interval, much wider than the final decimal places suggest.
Correction: Report the plug-in dose with units, then describe the uncertainty in the inputs, the propagation method, and the resulting interval or scenarios. Explain which sources of uncertainty were omitted. A recommendation is more defensible when the uncertainty range is interpreted against the decision, rather than when the point estimate is presented with unwarranted numerical precision.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If the point estimates in a dose equation are known to more decimal places, does that alone make the recommendation more reliable? Why?
- In the model \(CL=CL_{\mathrm{ref}}(W/70)^\theta M\), which input has log-clearance sensitivity \(\ln(W/70)\)?
- A linear renal-clearance model has slope 0.04 L/h per mL/min, and kidney function has standard error 5 mL/min. What is the propagated standard error of clearance if all other inputs are fixed?
- When both weight and maturation increase clearance, what does positive covariance between their log estimates do to the variance contribution?
- Why might a simulation-based interval still be misleading?
Solutions. (1) No. Extra decimal places describe the calculation’s numerical precision, not uncertainty in its inputs or the suitability of its model. (2) The allometric exponent \(\theta\), because differentiating \(\ln(CL)\) with respect to \(\theta\) gives \(\ln(W/70)\). (3) The propagated clearance standard error is \(0.04(5)=0.20\) L/h. (4) It adds a positive covariance contribution \(2\theta\operatorname{Cov}(\ln W,\ln M)\), increasing the variance when \(\theta\) is positive. (5) Simulation can precisely reflect unsuitable input distributions, omitted covariance, or a misspecified clearance model; computational detail does not make its assumptions valid.