Why the Source of Uncertainty Matters
A dose recommendation is based on quantities that are not known exactly. But “uncertainty” can refer to different things: uncertainty about a fitted dose-response curve, genuine variation in responses among patients, measurement or unexplained error, or uncertainty about whether the chosen model represents the response process adequately. These sources are not interchangeable. A calculation intended to estimate the population-average response at a dose does not have the same uncertainty target as a prediction for one future patient.
The practical question is therefore not simply how much uncertainty exists. It is which quantity is uncertain, and for whom or what is the decision being made? Separating the sources prevents a common error: adding every available variance component to every dose calculation, even when some components do not describe uncertainty in the decision target.
A Layered Dose-Response Model
Consider a response \(Y\) measured at dose \(d\). A useful general representation is \(Y=f(d,\theta,M)+b(d)+\varepsilon\). Here \(f\) is the systematic dose-response function, \(\theta\) denotes its parameters, and \(M\) denotes the chosen model structure, such as an Emax form or a linear form. The term \(b(d)\) represents between-subject deviation from the population response, while \(\varepsilon\) represents residual variation not explained by the model or subject effect. In applications, the subject effect may enter the function rather than be added to it; the distinction between sources remains the same.
Parameter uncertainty is an inferential statement about what the data tell us about fixed, unknown quantities; it does not require treating those quantities as physically random. A Bayesian analysis may represent it with a posterior distribution for \(\theta\), while a frequentist analysis may represent it using a sampling distribution or an approximation to the estimator’s uncertainty. Either way, it is conditional on a specified model unless model uncertainty is explicitly included.
Between-subject and residual variability describe response variation, not uncertainty about the model parameters. In a simple additive model, one might assume \(b\) has mean zero and variance \(\omega^2\), and \(\varepsilon\) has mean zero and variance \(\sigma^2\). If the two are independent conditional on the model parameters, their variances add for an individual response. A population mean averages over the distribution of \(b\); a prediction for one patient does not.
Structural uncertainty is different again. It arises when more than one scientifically plausible model or assumption could describe the data. A set of alternatives can be used for sensitivity analysis without assigning probabilities to them. It is only appropriate to average across model structures as if they had probabilities when defensible model weights have been specified.
What Should Enter the Calculation?
Start by defining the dose-related quantity used for the decision. For example, it might be the population mean response at dose \(d\), the response of a future patient at \(d\), or a dose that achieves a stated target. Then trace the sources of uncertainty that affect that quantity.
| Decision quantity | Sources that directly describe its uncertainty | Important qualification |
|---|---|---|
| Population mean response at a fixed dose, under one model | Parameter uncertainty in the population mean function | Between-subject and residual variation affect the data and thus parameter estimation, but are averaged out of the population mean itself. |
| Response of a future individual at a fixed dose | Parameter uncertainty, between-subject variability, and residual variability | Include components only as appropriate to the response definition and observation time. |
| Dose recommendation under alternative plausible model forms | Parameter uncertainty within each model, plus structural sensitivity across models | Do not treat model differences as a probability distribution unless model weights are justified. |
“Directly” is important. Between-subject and residual variability can be crucial to estimating \(\theta\): noisy or heterogeneous observations may make the fitted population curve less precise. Their influence is then reflected in parameter uncertainty. But that does not mean their full outcome variances should be added again to uncertainty in the population mean. Conversely, if the decision concerns a future individual response, the individual-level variability is part of the target and should not be omitted.
A Justification for the Variance Decomposition
The law of total variance explains how the layers combine when they are represented probabilistically. For any random variable \(Y\) and conditioning information \(Z\), it states that total variance is the variance of the conditional mean plus the expected conditional variance. Apply it first to the subject effect and residual error, conditional on parameters and model structure. Then account for uncertainty in the parameters.
For an individual response, the second term includes the conditional between-subject and residual variance. If those terms are independent, additive, and have variances \(\omega^2\) and \(\sigma^2\), respectively, that conditional variance is \(\omega^2+\sigma^2\). For a population mean response, the expectation over subjects removes a mean-zero subject effect, and residual observation error is not part of the mean estimand. The uncertainty in estimating that mean remains in the first term.
This derivation relies on the stated conditioning and model assumptions. In particular, the variance components must describe the relevant response scale and observation, and any dependence between components must be handled rather than silently ignored. If the residual variance changes with dose or the subject effect interacts with dose, a single constant variance may not be adequate.
Worked Example: Population Mean Versus One Patient
Suppose, hypothetically, that at a fixed dose the fitted population mean response has parameter uncertainty with variance \(4\) response-units squared. At that dose, between-subject variance is \(9\) response-units squared and residual variance is \(16\) response-units squared. Assume the additive components are independent and correctly specified.
For the population mean. Its standard uncertainty from the parameter layer is \(\sqrt{4}=2\) response units. The between-subject and residual variances are not added to the variance of the population mean itself: those components describe individual responses around that mean.
For one future individual response. Under the stated assumptions, the total variance is \(\operatorname{Var}(Y_{\mathrm{new}})=4+9+16=29\) response-units squared. Its standard deviation is \(\sqrt{29}=5.385\) response units. The value \(4\) is included because the population mean is estimated, while \(9\) and \(16\) describe subject-to-subject and residual response variation.
Thus, the same fitted model supports different uncertainty calculations depending on the target. Adding all three components is appropriate for this individual-response target, but not for uncertainty in the population mean alone.
Parameter Uncertainty in a Dose-Response Mean
When the target is the population mean at a fixed dose, parameter uncertainty depends on how the model function changes with its parameters. A simple linear model makes the role of parameter covariance visible. Let \(f(d,\theta)=\theta_0+\theta_1d\). If the estimated parameter vector has covariance matrix \(\Sigma\), then the variance of the estimated mean at fixed \(d\) is the variance of the linear combination \(\theta_0+d\theta_1\). The covariance term matters; ignoring it can overstate or understate uncertainty.
Worked Example: Propagating Parameter Covariance at a Fixed Dose
In a hypothetical linear dose-response analysis, the estimated intercept is \(10\) response units and the estimated slope is \(2\) response units per dose unit. At dose \(d=4\) dose units, suppose the parameter variances are \(\operatorname{Var}(\hat{\theta}_0)=0.25\), \(\operatorname{Var}(\hat{\theta}_1)=0.04\), and their covariance is \(-0.06\). Assume the linear form is the chosen model for this calculation.
The estimated population mean is \(\hat f(4)=10+(2)(4)=18\) response units. Its parameter variance is \(\operatorname{Var}(\hat{\theta}_0+4\hat{\theta}_1)=\operatorname{Var}(\hat{\theta}_0)+16\operatorname{Var}(\hat{\theta}_1)+8\operatorname{Cov}(\hat{\theta}_0,\hat{\theta}_1)\).
Substituting the values gives \(0.25+16(0.04)+8(-0.06)=0.25+0.64-0.48=0.41\) response-units squared. The corresponding standard uncertainty is \(\sqrt{0.41}=0.6403\) response units.
This calculation concerns uncertainty in the estimated population mean, not the spread of individual responses. If the hypothetical between-subject and residual variances at this dose were \(1\) and \(4\) response-units squared, respectively, the variance for a future individual response would instead be \(0.41+1+4=5.41\), with standard deviation \(\sqrt{5.41}=2.326\) response units. This addition assumes those components apply to the same individual response and are independent of each other conditional on the parameters.
Structural Uncertainty and Dose Recommendations
A dose recommendation can be sensitive to the model structure even when each model’s parameters are estimated precisely. This is especially important when the recommended dose is obtained by solving for a target response or optimizing an objective: alternative curves can imply different doses. Structural uncertainty should therefore be examined at the level of the decision, not only by comparing model fit summaries.
Worked Example: Two Plausible Curves, Two Target Doses
Suppose, hypothetically, two dose-response forms fit a small data set plausibly. Model A is \(f_A(d)=10+2d\), while Model B is \(f_B(d)=10+\frac{12d}{3+d}\), with dose \(d\) measured in dose units and response in response units. The target population mean response is \(18\) response units. Assume both curves are used only over a dose range where their forms are considered plausible.
Under Model A, set \(10+2d=18\). Then \(2d=8\), so the target dose is \(d=4\) dose units.
Under Model B, set \(10+\frac{12d}{3+d}=18\). Then \(\frac{12d}{3+d}=8\), so \(12d=24+8d\), \(4d=24\), and \(d=6\) dose units. At the Model A target dose of \(4\), Model B predicts \(10+\frac{48}{7}=16.86\) response units, not \(18\).
The two target doses differ by \(2\) dose units. This difference is evidence of structural sensitivity for the recommendation; it is not automatically a standard deviation or a probability statement. Without defensible probabilities for the models, report and investigate the model-specific results rather than averaging them as though the alternatives were random draws.
Diagnostics, Comparisons, and Common Pitfalls
Useful diagnostics ask whether each uncertainty component is supported by the data and relevant to the decision. Check residual patterns across dose and time; systematic patterns can indicate an inadequate mean function or residual model. Examine whether estimated between-subject variability is distinguishable from residual variability under the study design. Compare scientifically plausible structural alternatives and note whether they materially change the target dose or predicted response. A narrow parameter uncertainty measure within one model does not rule out substantial structural uncertainty.
The main alternative to a layered calculation is to use a single overall prediction or uncertainty measure. That can be convenient when the sole objective is prediction for a future observation and all relevant components are included. It is less informative for a dose decision because it can hide whether the uncertainty comes from limited information, patient heterogeneity, noisy measurement, or a fragile model choice. A sensitivity analysis across model structures is often clearer than forcing structural uncertainty into one number.
- Adding every variance component to a population-mean target. Between-subject and residual variation describe individual outcomes; they are not automatically variance in the population mean. They can affect estimation precision, which is already reflected in parameter uncertainty.
- Leaving variability out of a future-patient prediction. Parameter uncertainty alone describes uncertainty in the estimated mean function, not the full spread of future individual outcomes.
- Double counting. Do not add observation variability again if the quantity being combined is already a predictive distribution that includes it.
- Calling model spread a probability distribution. Results from several plausible models show sensitivity. They are not model probabilities unless weights have been assigned and justified.
- Treating all components as constant and independent. Dose-dependent variance, correlated effects, or nonlinear random effects may invalidate simple variance addition. State the assumptions and use a model that represents departures when needed.
Check Your Understanding
For each question, identify the target before deciding which uncertainty sources belong in the calculation. Solutions follow the questions.
- A decision concerns the population mean response at one fixed dose under a prespecified model. Which components directly describe uncertainty in that mean?
- A clinician wants the distribution of observed responses for a future patient at a fixed dose. Which components should be considered?
- Two plausible models give target doses of \(5\) and \(7\) dose units, but no model weights are available. What can be concluded, and what cannot?
- In a linear model, the slope estimate has variance \(0.09\), and the dose is \(3\) dose units. If the intercept is known exactly and there is no parameter covariance to include, what is the parameter variance of the estimated mean attributable to the slope?
- Why can it be wrong to add residual variance to parameter uncertainty for a population-mean dose target?
Solutions. (1) Parameter uncertainty in the population mean function directly describes uncertainty in the mean. Between-subject and residual variation may affect estimation but are not themselves variation in that mean. (2) Include parameter uncertainty, between-subject variability, and residual variability, subject to the model’s assumptions and the definition of the observed response. (3) The recommendation is structurally sensitive across these alternatives; the target-dose results can be reported model by model. Without weights, one cannot interpret their spread as a probability distribution over target doses. (4) The mean contribution is \(3\hat{\theta}_1\), so its variance is \(3^2(0.09)=0.81\) in squared response units. (5) Residual variance describes observation-to-observation deviations, not uncertainty in the population mean itself. Adding it can answer a different question—such as prediction for an individual—and can double count variability if it is already included elsewhere.