Why an Estimated Exposure Is Not a Direct Measurement
An exposure-response analysis often relates an individual-level PK summary, such as average concentration or area under the concentration-time curve, to an efficacy or safety outcome. That summary may be estimated from sparse concentration samples rather than observed directly. Assay error, limited sampling, and population-model assumptions can therefore make an individual exposure estimate differ from the person’s underlying exposure. If the estimated exposure is then treated as exact, the exposure-response relationship can be distorted.
The earlier material on propagating pharmacokinetic uncertainty into exposure-response establishes why PK uncertainty should not disappear when moving from a PK model to an outcome analysis. Here the focus is on how particular sources of error affect an individual exposure predictor, what can and cannot be concluded from an empirical Bayes estimate, and how regression calibration and joint modeling handle uncertainty.
These sources are related but not interchangeable. Assay error occurs at the concentration-measurement level. Sparse sampling leaves less information for estimating an individual’s PK parameters and derived exposure. A population PK model combines individual data with population information, so its EB estimates may be less variable than the underlying individual values. The resulting exposure-estimation error need not be independent of the true exposure, constant across individuals, or even well represented by a single additive-error model.
Classical Error and Attenuation in a Linear Model
A useful reference case is a linear outcome model with classical measurement error. Suppose the true exposure-response relationship is \(Y_i=\alpha+\beta X_i+\varepsilon_i\), but the analysis uses \(W_i=X_i+u_i\). Assume that \(u_i\) has mean zero, is independent of \(X_i\) and \(\varepsilon_i\), and has finite variance; also assume \(E(\varepsilon_i\mid X_i,u_i)=0\). Under these assumptions, regressing \(Y_i\) on \(W_i\) does not, in general, estimate \(\beta\).
The result follows from the population regression slope, \(\mathrm{Cov}(W,Y)/\mathrm{Var}(W)\). Independence and zero-mean error give \(\mathrm{Cov}(W,Y)=\mathrm{Cov}(X+u,\alpha+\beta X+\varepsilon)=\beta\mathrm{Var}(X)\), while \(\mathrm{Var}(W)=\mathrm{Var}(X)+\mathrm{Var}(u)\). Dividing these quantities gives the stated result. This derivation is specific to the assumptions: it is not a general rule that every kind of exposure-estimation error attenuates every exposure-response estimate.
Worked Example: Classical Error Attenuates a Linear Slope
Suppose a hypothetical study has a true exposure variance of \(100\;(\mathrm{mg/L})^2\), an independent exposure-estimation error variance of \(25\;(\mathrm{mg/L})^2\), and a true linear outcome slope of \(0.40\) outcome points per \(\mathrm{mg/L}\). Under the classical-error assumptions, the reliability ratio is
The expected plug-in slope is therefore
In this hypothetical setting, using the noisy exposure estimate is expected to understate the positive slope by \(0.080\) outcome points per \(\mathrm{mg/L}\). This calculation is justified only if the additive error is independent of true exposure and outcome residuals, and if the stated linear model is adequate. It is not a correction to apply automatically to EB estimates.
Assay error may be approximately additive on the concentration scale, while proportional error is often more plausible on a logarithmic scale. Repeated assay measurements can help quantify assay variability, but they do not by themselves resolve uncertainty caused by sparse sampling or an inadequate PK model. A more complete analysis distinguishes the concentration observation model from the PK model that maps concentrations to an exposure summary.
What Shrinkage Does—and Does Not—Tell You
An EB estimate uses the individual’s PK data together with population information. When individual data are sparse or weakly informative, the estimate is pulled toward the population prediction; this is shrinkage. The pull is often stronger for individuals with less informative data. Consequently, the distribution of EB estimates can have less between-individual spread than the distribution of true exposures.
Reduced spread can matter substantially when the estimated exposure is used as if it were the true predictor. But shrinkage alone does not determine the direction or size of bias in an exposure-response coefficient. EB estimates are not necessarily the classical-error variable \(X+u\): their errors can depend on the data, covariates, model structure, and degree of individual information. In some settings, treating compressed EB estimates as exact can exaggerate a slope; in others, uncertainty can attenuate or otherwise distort it. A linear outcome model combined with a correctly calibrated conditional mean of exposure has special properties that can avoid slope bias. Those properties should not be assumed merely because an estimate is labelled EB.
Worked Example: A Compressed Predictor Can Change a Plug-In Slope
Consider a small, deliberately simplified hypothetical illustration. The true exposures for two individuals are \(8\) and \(12\;\mathrm{mg/L}\), and their outcomes follow \(Y=20+3(X-10)\), giving \(14\) and \(26\) outcome points. Suppose the exposure values entered in a naïve regression are instead compressed toward \(10\;\mathrm{mg/L}\) by the rule \(W=10+0.40(X-10)\). The entered values are \(9.2\) and \(10.8\;\mathrm{mg/L}\).
The two-point regression slope is the difference in outcomes divided by the difference in entered exposures:
The true slope is \(3.00\) outcome points per \(\mathrm{mg/L}\), whereas this compressed-predictor calculation gives \(7.50\). This is an algebraic illustration of what can happen when a predictor is compressed but the corresponding individual-level uncertainty is omitted; it is not a general prediction for fitted EB estimates. If the EB value is a well-calibrated conditional mean and the outcome model is linear with suitable error assumptions, the conditional mean of the outcome given that estimate may instead retain the true slope. The example’s lesson is to model or assess the uncertainty structure, not to apply a universal shrinkage correction.
Regression Calibration
Regression calibration replaces an uncertain exposure with its conditional expectation given the available exposure information and relevant covariates. Write \(m(W_i,C_i)=E(X_i\mid W_i,C_i)\), where \(C_i\) denotes covariates used in the calibration model. For a linear outcome model, if \(E(Y_i\mid X_i,C_i)=\alpha+\beta X_i+\gamma^\mathsf{T}C_i\), then
provided the outcome residual has conditional mean zero given \(X_i,W_i,C_i\). This identity justifies using the calibrated exposure in the conditional mean for this linear setting. It does not mean the latent exposure has been observed, and it does not remove all uncertainty from estimates or predictions.
Worked Example: Calibrating a Noisy Exposure Estimate
Assume, hypothetically, that \(X\sim N(10,4)\;(\mathrm{mg/L})\), and an exposure estimate \(W=X+u\) has independent error \(u\sim N(0,5)\;(\mathrm{mg/L})\). For this normal model, the conditional mean is the prior mean plus a variance-weighted adjustment:
For an individual with \(W=13\;\mathrm{mg/L}\), the calibrated exposure is
If the hypothetical conditional outcome mean is \(E(Y\mid X)=50+1.5X\) outcome points, the calibrated conditional mean is \(50+1.5(11.33)=67.0\) outcome points. Direct substitution of \(W=13\) would give \(69.5\) points. This calibration is appropriate only to the extent that the normal exposure and error model, independence assumption, and linear outcome mean are reasonable. In an analysis, calibration parameters are estimated, so their uncertainty should also be considered.
Regression calibration is attractive when an exposure measurement model can be fitted separately and the outcome relationship has a form for which replacing \(X\) by \(E(X\mid W,C)\) is justified or is an adequate approximation. The simple conditional-mean identity above is exact for the specified linear conditional mean. For a nonlinear outcome mean, in general \(E\{f(X)\mid W,C\}\ne f\{E(X\mid W,C)\}\); substituting a calibrated mean into the nonlinear function can therefore remain biased. Calibration also relies on having enough information to estimate the conditional exposure distribution and on accounting appropriately for covariates that predict both exposure and outcome.
Joint Models and Uncertainty Propagation
A joint model treats exposure as latent and combines a PK measurement model with an outcome model. A schematic individual likelihood contribution is
where \(D_{\mathrm{PK},i}\) denotes the individual’s concentration data. In a mechanistic PK analysis, the PK component may instead be expressed through individual PK parameters and a derived exposure \(X_i\). The essential point is that the likelihood accounts for the distribution of plausible exposures rather than fixing exposure at one estimate. A joint analysis can propagate sparse-sampling and assay uncertainty into exposure-response estimates and can allow outcome data to contribute information about latent exposure while estimating the outcome relationship.
Worked Example: Outcome Data Update a Latent Exposure in a Joint Model
For a hypothetical illustration, suppose \(X\sim N(10,4)\;\mathrm{mg/L}^2\) in the sense that the mean is \(10\;\mathrm{mg/L}\) and the variance is \(4\;(\mathrm{mg/L})^2\). A PK-derived estimate is \(W=13\;\mathrm{mg/L}\) with error variance \(5\;(\mathrm{mg/L})^2\). Suppose the outcome is \(R=68\) points and, conditional on \(X\), \(R\sim N(50+1.5X,9)\), with outcome variance \(9\) points squared. Treat all model parameters and variances as known for this single-individual calculation.
The normal posterior precision is the sum of prior, PK, and outcome precisions. The outcome contributes precision \((1.5)^2/9=0.25\;(\mathrm{mg/L})^{-2}\), because its mean changes by \(1.5\) points per \(\mathrm{mg/L}\). Thus
Using PK data alone would give \(11.33\;\mathrm{mg/L}\), as in the preceding example; the outcome shifts the joint posterior mean to \(11.57\;\mathrm{mg/L}\). This illustrates joint information use, not a recipe for estimating a population exposure-response slope from one individual. In a study, the PK and outcome models, population parameters, and dependence assumptions must be estimated and checked together. Outcome-informed exposure estimates should not then be treated as independent predictors in a second analysis, which would risk using the outcome information twice.
Diagnostics, Alternatives, and Limitations
No single diagnostic proves that exposure estimates are reliable, but several checks can reveal problems. Examine the informativeness of individual PK data and the amount of EB shrinkage, including whether shrinkage differs by dose, sampling design, or covariates. Compare the spread and ranking of individual exposure estimates with model-based uncertainty or simulation-based predictions. Assess assay precision and the consequences of plausible assay-error models. If the exposure-response conclusion changes materially when uncertainty is propagated, the plug-in analysis is not robust to exposure estimation.
The main alternatives address different needs:
| Approach | What it uses | Main caution |
|---|---|---|
| Plug-in regression | One exposure estimate per individual, treated as exact | Can ignore uncertainty and distort the slope; convenient but often insufficient when estimates are imprecise. |
| Regression calibration | Conditional mean of latent exposure given exposure data and covariates | Depends on a credible calibration model and is not generally exact for nonlinear outcome means. |
| Joint model | PK data, latent exposure distribution, and outcome likelihood together | Requires a coherent joint specification and adequate computation; misspecification in either component can affect inference. |
These methods do not repair a poorly specified PK model, remove confounding, or establish that an exposure-response association is causal. They address uncertainty in the exposure representation under their modeling assumptions. Nor does a small estimated standard error from a plug-in regression show that exposure uncertainty is negligible: that standard error is conditional on the exposure values supplied to the regression.
The practical decision is not simply whether to use an EB estimate or a more complex method. It is whether the uncertainty and dependence relevant to the scientific question are adequately represented. When exposure estimates are precise and conclusions are stable, a carefully qualified plug-in analysis may be serviceable. With sparse data, substantial shrinkage, or conclusions sensitive to exposure uncertainty, calibration or joint modeling is more defensible. Each approach remains limited by its assumptions and by the information in the study design.
Check Your Understanding
Try each question, then compare your reasoning with the brief solutions.
- Under classical additive error, what happens to a positive linear slope when error variance increases while true-exposure variance stays fixed?
- Why is assay error not the same as uncertainty in an AUC estimated from sparse samples?
- Does EB shrinkage alone establish that a fitted exposure-response slope is attenuated? Explain.
- Under what outcome-model condition does regression calibration have the simple conditional-mean justification given here?
- What is one advantage and one limitation of a joint model compared with plug-in regression?
Solutions. (1) The reliability ratio \(\mathrm{Var}(X)/[\mathrm{Var}(X)+\mathrm{Var}(u)]\) decreases, so the expected slope moves closer to zero under the classical-error assumptions. (2) Assay error concerns measured concentrations; sparse sampling also limits estimation of the PK profile and its derived AUC, even with a precise assay. (3) No. EB errors can be structured, and slope bias depends on their relationship to true exposure and the outcome model. (4) The conditional outcome mean must be linear in exposure, with an outcome residual having conditional mean zero given the exposure information and covariates. (5) A joint model can propagate latent-exposure uncertainty, but it requires a coherent and adequately specified PK and outcome model.