Why Uncertainty Must Be Traced Through a Model Chain
A downstream prediction often depends on quantities estimated or assumed upstream. For example, a predicted clinical response may use an exposure estimate from a pharmacokinetic model and a parameter from an exposure-response model. Reporting only the uncertainty from the final model can give an incomplete picture if upstream uncertainty was treated as known, or if dependencies between shared parameters were ignored.
The review question is not simply whether a prediction has a standard error or interval. It is which sources of uncertainty were carried into that result, and which were held fixed or omitted? A clear report identifies the chain of inputs, states how uncertainty was represented at each link, and explains what the final interval or distribution describes.
Keep distinct the uncertainty in an estimated model quantity and variability among individual outcomes. Parameter or input uncertainty concerns what is not known about a model quantity, such as a population exposure estimate. Residual or individual-level variability concerns variation in outcomes not captured by the modeled predictors. A prediction interval intended to describe individual outcomes may need both; uncertainty in a population mean prediction does not, by itself, describe the spread of individual responses.
Basic Results for Linked Quantities
Let \(Z\) denote uncertain upstream quantities and let the downstream quantity be \(Y=g(Z)\). If the model treats \(Z\) as random under a stated uncertainty distribution, the law of total variance separates variation in conditional means from variation remaining within those conditions:
The first term reflects how the expected downstream result changes as upstream quantities vary. The second term reflects conditional variability in \(Y\) after \(Z\) is specified, such as residual variability if the prediction targets individual outcomes. If \(Y\) is a deterministic model output once \(Z\) is given, the conditional variance is zero. This decomposition is useful only if the random quantities and their distributions represent the uncertainty relevant to the question.
For a smooth function \(g\), a common local approximation uses a first-order Taylor expansion around the mean vector \(\boldsymbol{\mu}\) of \(Z\). If \(Z\) has covariance matrix \(\boldsymbol{\Sigma}\), then:
The approximation follows because the constant \(g(\boldsymbol{\mu})\) contributes no variance, and the variance of a linear combination \(a^\mathsf{T}Z\) is \(a^\mathsf{T}\boldsymbol{\Sigma}a\). For two uncertain inputs \(Z_1,Z_2\), this becomes:
with derivatives evaluated at the chosen expansion point. The covariance term matters when inputs are dependent. The approximation can be poor when the function is strongly nonlinear over the relevant uncertainty range, the distribution is skewed, or the inputs have substantial probability near a boundary.
Worked Example: Carrying Exposure and Slope Uncertainty Into a Response Prediction
Suppose a hypothetical linked analysis predicts a response measure \(Y=0.4+\beta E\), where exposure \(E\) has mean 100 exposure units and standard error 10 exposure units, and the response slope \(\beta\) has mean 0.015 response units per exposure unit and standard error 0.002 response units per exposure unit. For this calculation, assume the two estimates are independent, the intercept 0.4 is known exactly, and the target is uncertainty in the predicted mean, not individual residual variability.
The plug-in prediction is:
For independent uncertain \(E\) and \(\beta\), the variance of their product is \(\operatorname{Var}(\beta E)=\mu_\beta^2\operatorname{Var}(E)+\mu_E^2\operatorname{Var}(\beta)+\operatorname{Var}(\beta)\operatorname{Var}(E)\). Here \(\operatorname{Var}(E)=10^2=100\) exposure-units squared and \(\operatorname{Var}(\beta)=0.002^2=0.000004\) response-units squared per exposure-unit squared. Thus:
A first-order normal approximation gives \(1.90\pm1.96(0.2508)\), or approximately 1.408 to 2.392 response units. This is an approximate 95% uncertainty interval under the stated input distributions and independence assumption; it is not a statement that the fixed predicted value has a 95% probability of lying in this particular interval. If exposure were treated as known, its variance contribution and the product term would be excluded, yielding \(SE(Y)=100(0.002)=0.200\) response units. That smaller result answers a conditional question: uncertainty in the slope while exposure is held fixed.
Shared Parameters and Dependence
Linked models may use the same estimated parameter, or parameter estimates may be correlated because they came from a joint fit. In either case, carrying each marginal standard error forward while assuming independence can distort the downstream uncertainty. Positive covariance increases the variance of a sum when both quantities contribute in the same direction; negative covariance can reduce it. For a difference, the direction of the effect reverses.
Worked Example: Retaining Covariance in a Downstream Contrast
Consider a hypothetical downstream contrast \(M=A-B\), where \(A\) and \(B\) are two jointly estimated components in compatible units. Their estimates are \(A=2.0\) units with \(SE(A)=0.30\) units and \(B=1.0\) unit with \(SE(B)=0.20\) units. Their estimated correlation is 0.80. Assume the reported covariance matrix is an appropriate representation of parameter uncertainty for this calculation.
The estimated covariance is \(\operatorname{Cov}(A,B)=0.80(0.30)(0.20)=0.048\) units squared. The contrast estimate is \(M=2.0-1.0=1.0\) unit, and its variance is:
Ignoring covariance would give \(SE(M)=\sqrt{0.090+0.040}=0.3606\) units, almost twice as large. Here the positive covariance reduces uncertainty in the difference because shared movement in \(A\) and \(B\) cancels. A normal approximation gives a 95% interval of \(1.0\pm1.96(0.1844)\), approximately 0.639 to 1.361 units. The result depends on the covariance being retained and on the normal approximation being reasonable; marginal standard errors alone do not reproduce this calculation.
Fixed Inputs and Omitted Sources
An input held fixed is not necessarily known without error. It may be fixed because the analysis conditions on a specified scenario, because its uncertainty is judged negligible, or simply because no uncertainty distribution was supplied. These rationales have different implications and should not be conflated. A reviewer should be able to tell whether a reported interval is conditional on a dose, covariate, external estimate, or structural assumption being exactly as specified.
Worked Example: Quantifying the Effect of a Fixed Assumption
Suppose a hypothetical downstream efficacy index is \(Y=FE\), where an upstream model gives exposure \(E\) with mean 100 units and standard error 10 units. A conversion factor \(F\), representing an assumed transfer from exposure to the index, is assigned the value 0.80. First treat \(F\) as fixed. Then consider a sensitivity calculation in which \(F\) has mean 0.80 and standard error 0.10, independently of \(E\). These are illustrative uncertainty distributions, not observed study results.
With \(F\) fixed, the prediction is \(0.80(100)=80.0\) index units and \(SE(Y)=0.80(10)=8.00\) index units. When uncertainty in \(F\) is represented, the product-variance formula gives:
The mean remains 80.0 index units under independence, but the propagated standard error increases from 8.00 to 12.85 when the conversion factor is uncertain. This comparison shows what the fixed-input result leaves out; it does not establish that the illustrative distribution for \(F\) is scientifically appropriate. The report should state the basis for fixing \(F\), and, if material, show how plausible alternatives affect the downstream conclusion.
Reviewing and Diagnosing Propagation
For nonlinear or complex linked models, simulation can propagate uncertainty without relying on a local linear approximation. A typical calculation draws parameter and input values from their joint uncertainty distribution, passes each draw through every model in sequence, and summarizes the resulting downstream predictions. The draws must preserve dependencies: independently sampling parameters that were estimated jointly discards covariance. If the target includes individual outcomes, the simulation must also represent the relevant residual or individual-level variability, rather than reporting only variation in predicted means.
A practical review can follow this sequence:
List each upstream estimate, fixed input, assumption, and downstream transformation that contributes to the reported quantity.
For every quantity, determine whether it is treated as fixed, assigned a distribution, or represented through a covariance matrix or joint draws.
Confirm that shared parameters and correlated inputs remain dependent, and that the uncertainty measure matches the target: for example, a population mean or an individual outcome.
Where practical, compare an analytic approximation with propagated draws or relevant alternative assumptions, then state what sources remain excluded.
Warning signs include a downstream interval narrower than expected because upstream estimates were silently fixed; a simulation that samples correlated parameters independently; an interval described as individual-level prediction when it includes parameter uncertainty only; or a normal interval used despite a strongly skewed or bounded output. For a first-order approximation, compare the linear prediction with evaluations over plausible input ranges. For simulation, inspect whether draws are valid, whether extreme values dominate, and whether increasing the number of draws materially changes reported summaries. These checks assess the propagation calculation; they do not validate the upstream models or make omitted assumptions disappear.
Choosing an Approach and Stating Limitations
| Approach | Useful when | Key limitation to disclose |
|---|---|---|
| First-order variance approximation | The downstream function is smooth and uncertainty is modest near the expansion point. | May miss curvature, skewness, or boundary effects; results depend on the covariance inputs. |
| Joint simulation through linked models | The chain is nonlinear or distributions of downstream predictions are needed. | Depends on the chosen input distributions, preserved dependencies, and adequate simulation coverage. |
| Fixed-input calculation | The analysis intentionally conditions on a specified value or scenario. | Does not include uncertainty in that input; interpretation must remain conditional. |
No propagation method can compensate for an omitted source that was never represented, an inappropriate uncertainty distribution, or a misspecified model relationship. Local approximations are efficient but can be unreliable for nonlinear transformations. Simulation can represent complex transformations, but its output is only as defensible as its joint input distributions and model chain. Scenario analyses can expose dependence on assumptions, but a set of scenarios is not automatically a probability distribution over those assumptions.
A transparent report names the downstream target and units; describes upstream quantities and their uncertainty; explains whether shared parameters and covariance were retained; identifies inputs held fixed and why; and distinguishes parameter uncertainty from residual or individual variability. When some sources are omitted, state that limitation and avoid presenting the resulting interval as unconditional or comprehensive.
Common Pitfall
Wrong reasoning: “The downstream model reports a narrow interval, so uncertainty from the upstream model must already be included.” A downstream model may have been evaluated at a single upstream point estimate, in which case its interval is conditional on that value. Correction: trace each uncertain quantity through the calculation and verify that its variance, distribution, or joint dependence was actually propagated. If it was fixed, describe the result as conditional and assess whether that conditioning is appropriate for the claim.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A downstream prediction uses an upstream estimate with standard error 5 units, but the report says the upstream value was fixed at its estimate. What uncertainty is absent from the reported downstream calculation?
- Two jointly estimated quantities have positive covariance and enter a downstream result as their difference. What is the covariance term in the variance, and how does positive covariance affect it?
- Why might a first-order variance approximation be unreliable for a strongly nonlinear transformation?
- A simulation draws each parameter independently despite a reported joint covariance matrix. What dependency has been lost, and why can the resulting uncertainty be wrong?
- What additional variation is needed if the target changes from uncertainty in a population mean prediction to a prediction for an individual outcome?
Solutions. (1) The contribution of uncertainty in the upstream estimate is absent; the reported result is conditional on the fixed value. (2) For a difference \(A-B\), the term is \(-2\operatorname{Cov}(A,B)\); positive covariance reduces the variance because shared movement cancels. (3) The local linearization may not represent curvature across the input range, so it can misstate the output variance or distribution. (4) The simulation loses parameter dependence and omits covariance contributions, which can inflate or reduce downstream uncertainty depending on the model relationships. (5) Include relevant residual or individual-level variability as well as parameter uncertainty, if the individual prediction is intended to represent both.