Why uncertainty errors matter
An uncertainty-aware dose recommendation depends not only on the fitted model, but also on how its uncertainty is represented and used. A calculation can be numerically correct and still answer the wrong question: treating an estimated parameter as fixed omits uncertainty about the model, while adding patient-to-patient variability to a parameter interval confuses two distinct sources of variation. Ignoring covariance can misstate uncertainty in a combined prediction, and extending a fitted curve far beyond observed doses can make a precise-looking recommendation rest on weak evidence.
These are consequential errors because dose choice compares predictions across doses. If an error changes the estimated probability of meeting a target or the expected value of a dose, it may change the selected dose. The goal is not to make every calculation maximally conservative. It is to match the calculation to the decision quantity, preserve the uncertainty structure, and make clear what the data do and do not support.
Let \(d\) denote dose, \(\theta\) the model parameters, and \(Y\) an outcome for a future patient or observation. A fitted model may specify the distribution of \(Y\) conditional on \(d\) and \(\theta\). An uncertainty-aware prediction also accounts for uncertainty in \(\theta\), using a posterior distribution or an appropriate sampling distribution. The right probability depends on the question. For example, the probability that a latent typical response exceeds a threshold is not the same as the probability that a future patient's observed response exceeds it.
Use the right uncertainty for the decision
A common plug-in calculation substitutes a point estimate \(\hat{\theta}\) into the model and proceeds as if that estimate were known exactly. This can be useful as a descriptive summary, but it does not generally provide a probability that integrates parameter uncertainty. A different common error is to treat all predictive spread as parameter uncertainty. The predictive distribution for a future outcome can include both uncertainty about \(\theta\) and variability in \(Y\) conditional on \(\theta\).
The law of total probability and the law of total variance make the distinction precise. Conditional probabilities are averaged over parameter uncertainty, and total predictive variance has a component from each source.
The first variance term represents conditional outcome variability averaged over parameter uncertainty; the second represents variation in the model's conditional mean as the uncertain parameters vary. These equations do not say to add every available variance mechanically. The model and the decision target determine which sources belong in the calculation. For a decision about a latent population mean, for example, residual variation among individual observations may not be part of the target quantity.
Worked Example: Separating a Future Patient's Variability From Parameter Uncertainty
Suppose a hypothetical model at a chosen dose has conditional outcome \(Y=\theta+\varepsilon\), where the uncertain mean response has distribution \(\theta\mid\text{data}\sim N(1.20,0.40^2)\), and a future patient's residual outcome satisfies \(\varepsilon\sim N(0,0.50^2)\), independently of \(\theta\). Response and threshold are measured in the same hypothetical response units. What is the predictive probability that the patient's response is at least 1.00 units?
If the point estimate \(\hat{\theta}=1.20\) is treated as known, the plug-in calculation is \(P(Y\geq1.00\mid\hat{\theta})=\Phi((1.20-1.00)/0.50)=\Phi(0.400)=0.6554\). This is conditional on a fixed mean response; it does not integrate uncertainty in that mean.
For the predictive probability, independence and normality give \(Y\mid\text{data}\sim N(1.20,0.40^2+0.50^2)=N(1.20,0.4100)\). Thus the predictive standard deviation is \(\sqrt{0.4100}=0.6403\) response units, and \(P(Y\geq1.00\mid\text{data})=\Phi((1.20-1.00)/0.6403)=\Phi(0.3123)\approx0.6226\).
The probability that the latent mean itself exceeds 1.00 is a third, different quantity: \(P(\theta\geq1.00\mid\text{data})=\Phi((1.20-1.00)/0.40)=\Phi(0.500)=0.6915\). The three probabilities answer different questions. For an individual patient's target attainment, use the predictive probability, not the probability about the latent mean. All values here are hypothetical.
Carry covariance into combined predictions
Model parameters are often estimated together, so their uncertainties need not be independent. If a decision quantity combines parameters—for example, a benefit prediction minus a harm prediction—its variance depends on their covariance. Dropping that covariance is not automatically conservative: positive covariance can reduce the variance of a difference, while negative covariance can increase it.
For two uncertain quantities \(A\) and \(B\), the variance of their difference is \(\operatorname{Var}(A-B)=\operatorname{Var}(A)+\operatorname{Var}(B)-2\operatorname{Cov}(A,B)\). This follows by expanding \(E[((A-E[A])-(B-E[B]))^2]\); the cross-product contributes \(-2\operatorname{Cov}(A,B)\). For a general linear combination \(a^\mathsf{T}\theta\), its variance is \(a^\mathsf{T}\Sigma a\), where \(\Sigma\) is the parameter covariance matrix. Both results assume finite variances and the stated joint distribution; normality is additionally needed if the resulting normal probability calculation is used.
Worked Example: The Effect of Covariance on a Benefit–Harm Margin
Suppose a hypothetical dose-related calculation expresses efficacy benefit and toxicity burden in common normalized units. Their uncertain coefficients are jointly normal, with \(E(\beta_E)=1.40\), \(SD(\beta_E)=0.30\), \(E(\beta_T)=0.90\), \(SD(\beta_T)=0.25\), and correlation \(0.80\). At the dose being assessed, the net margin is \(M=\beta_E-\beta_T\). Calculate the probability that the margin is positive.
The expected margin is \(E(M)=1.40-0.90=0.500\) normalized units. The covariance is \(\operatorname{Cov}(\beta_E,\beta_T)=0.80(0.30)(0.25)=0.0600\) squared normalized units. Therefore, \[ \operatorname{Var}(M)=0.30^2+0.25^2-2(0.0600)=0.03250, \] and \(SD(M)=\sqrt{0.03250}=0.1803\) normalized units. Under the joint-normal assumption, \(M\) is normal, so \(P(M>0)=\Phi(0.500/0.1803)=\Phi(2.773)\approx0.9972\).
If covariance were incorrectly set to zero, the variance would be \(0.30^2+0.25^2=0.1525\), the standard deviation would be \(0.3905\), and the calculated probability would be \(\Phi(0.500/0.3905)=\Phi(1.28041)\approx0.8998\). Here, omitting positive covariance overstates uncertainty in a difference because the two coefficients tend to move together. This example illustrates why the covariance matrix, not just separate standard errors, is needed for a joint prediction. The calculation is hypothetical.
Check whether the chosen dose is supported by the data
A fitted dose-response curve may produce predictions at doses not represented in the data. Such extrapolation is not inherently invalid: a mechanistic model, external evidence, or a justified structural assumption may support it. The error is to treat a distant prediction as equally well established as an interpolation, or to report a precise dose threshold without checking how strongly it depends on the model outside the observed range.
For an Emax model \(E(d)=E_{\max}d/(ED_{50}+d)\), the response approaches \(E_{\max}\) as dose increases. A point estimate can imply a threshold-crossing dose, but uncertainty in \(E_{\max}\) or \(ED_{50}\), misspecification of the curve, and sparse observations near the upper range may substantially change that crossing. Checking only the fitted curve's smoothness is not a validation of its extrapolation.
Worked Example: A Threshold Dose Beyond the Observed Range
In a hypothetical study, responses are observed only through 100 mg. A fitted Emax model has point estimates \(E_{\max}=100\) response units and \(ED_{50}=50\) mg. The decision target is a response of at least 75 units. Calculate the point-estimate threshold dose, then assess the extrapolation.
Using \(E(d)=100d/(50+d)\), the threshold equation is \(100d/(50+d)\geq75\). Since doses are nonnegative, \(50+d>0\), so multiplying by the denominator gives \(100d\geq3750+75d\), hence \(25d\geq3750\), or \(d\geq150\) mg. The point-estimate curve therefore first reaches the target at 150 mg, which is 50 mg above the largest observed dose.
At the largest observed dose, the fitted response is \(E(100)=100(100)/(50+100)=66.67\) units. At 150 mg it is exactly \(100(150)/(50+150)=75.00\) units, and at 200 mg it is \(100(200)/(50+200)=80.00\) units. A sensitivity check holding \(E_{\max}=100\) fixed but changing \(ED_{50}\) illustrates the dependence: at 200 mg, \(ED_{50}=30\) mg gives \(E(200)=100(200)/(30+200)=86.96\) units, whereas \(ED_{50}=70\) mg gives \(E(200)=100(200)/(70+200)=74.07\) units. These are illustrative parameter values, not a confidence interval or probability statement.
The calculation establishes what the point-estimate model predicts; it does not establish that 150 mg is a supported recommendation. Because that dose is outside the observed range, a decision would need to account for parameter and model uncertainty and justify extrapolation, for example with relevant external evidence or additional data. The example is hypothetical.
Diagnostics, alternatives, and limitations
A useful diagnostic begins with the decision quantity, not with the software output. State whether the target concerns a latent mean, a typical patient, or a future individual; identify which model parameters are uncertain; and identify which sources of outcome variability belong in the target. Then check whether the joint parameter uncertainty has been retained and whether the recommended dose lies within the domain supported by data or other evidence.
Write the target as a probability, expected utility, or constraint in context. Distinguish a statement about parameters from a prediction for a future patient.
Separate parameter uncertainty, residual variability, and any modeled population heterogeneity. Include each only when it belongs to the stated target.
For combined predictions, check that relevant covariances are retained and that the resulting variance or simulated distribution is plausible.
Compare the recommendation with observed doses, inspect influential observations and model fit, and evaluate how the recommendation changes under plausible parameter or structural alternatives.
Warning signs include a recommendation that changes sharply when a standard error or covariance is altered, a prediction interval that is implausibly narrow, or a dose threshold located far beyond the data with no sensitivity analysis. Simulation from a fitted joint parameter distribution can propagate nonlinear uncertainty more faithfully than a first-order approximation, but it cannot repair a misspecified model or unsupported extrapolation. A plug-in estimate is simpler and can be adequate for a clearly labeled point prediction; it is not a substitute for uncertainty integration when the decision rule depends on uncertainty. Sensitivity analysis explores dependence on assumptions, while additional data or external validation may be needed to resolve weak support.
Common mistakes
- Treating an estimate as known. The incorrect reasoning is that a fitted value is the true parameter because it is the model's best estimate. Instead, propagate its uncertainty when the decision concerns uncertain predictions or expected utility.
- Equating variability with uncertainty. The incorrect reasoning is that a broad distribution among patients means the fitted mean is uncertain by the same amount. Instead, distinguish conditional patient variability from uncertainty about model parameters; they answer different questions.
- Dropping covariance. The incorrect reasoning is that separate parameter standard errors fully determine uncertainty in a combined quantity. Instead, use the covariance terms implied by the joint model or fitted covariance matrix.
- Trusting an extrapolated threshold. The incorrect reasoning is that a smooth curve gives reliable predictions wherever the formula can be evaluated. Instead, check data support, sensitivity, and external justification beyond the observed range.
No single calculation eliminates these limitations. An uncertainty-aware recommendation remains conditional on the model, the available data, and the chosen decision target. Sensitivity to those choices should be investigated rather than hidden by a single point estimate.
Check Your Understanding
Use the definitions and calculations above to answer each question. Full solutions follow the questions.
- A model gives a conditional response standard deviation of 3 units and an estimated mean with standard error 4 units. Which quantity represents patient-level variability, and which represents uncertainty in the estimated mean?
- Two uncertain quantities have variances 0.16 and 0.09 and covariance 0.06. Find the variance of their difference. What would the variance be if covariance were incorrectly ignored?
- A fitted response curve reaches its target at 180 mg, but the largest observed dose was 120 mg. Is 180 mg automatically an invalid dose, a validated dose, or an extrapolated point estimate requiring further assessment?
- In a normal predictive model, why can the probability that the latent mean exceeds a threshold differ from the probability that a future patient's response exceeds it?
Solutions. (1) The conditional standard deviation of 3 units describes patient-level response variability under the model; the standard error of 4 units describes uncertainty in the estimated mean. (2) The difference variance is \(0.16+0.09-2(0.06)=0.13\). Ignoring covariance gives \(0.16+0.09=0.25\). (3) It is an extrapolated point estimate requiring further assessment. It is not automatically invalid, but the data through 120 mg alone do not validate the prediction at 180 mg. (4) The latent-mean probability concerns uncertainty in the mean parameter only, whereas a future patient's response also varies conditionally around that mean. The probabilities therefore refer to different random quantities and generally differ.