Why Estimated Exposure Is Not Error-Free
An exposure-response analysis often uses an individual exposure summary, such as area under the concentration-time curve, derived from a population PK model. That summary is an estimate, not a directly observed value. If it is treated as exact in the clinical outcome model, uncertainty in its estimation can distort the exposure-response relationship. In the simplest case, random error in exposure estimates pulls a linear slope toward zero. In practice, the error may also vary across individuals, depend on sampling, or be related to outcome predictors.
The relevant distinction is between an individual's latent exposure, which is the quantity the outcome model aims to relate to response, and the estimated exposure available for analysis. The methods below differ in how they represent the link between those quantities and how much information they use to estimate it.
A useful starting model is classical additive measurement error:
Here \(U_i\) is exposure error and \(\varepsilon_i\) is outcome-model error. For the basic derivation, assume \(X_i\), \(U_i\), and \(\varepsilon_i\) are mutually independent, have finite variances, and \(E[U_i]=E[\varepsilon_i]=0\). Also assume that the linear conditional mean \(E[Y_i\mid X_i]=\alpha+\beta X_i\) is appropriate. These assumptions define a teaching model, not a default description of every population PK estimate.
How Classical Error Attenuates a Slope
If \(Y_i\) is regressed directly on \(W_i\), the population regression slope is \(\operatorname{Cov}(W,Y)/\operatorname{Var}(W)\). Independence and the model above give:
Therefore, the slope from regressing outcome on the error-prone exposure is:
The factor \(\lambda\) is the reliability ratio under this model. The derivation shows why the slope is attenuated toward zero: the noisy predictor has more variance than the true exposure, but its covariance with the outcome is determined by the true-exposure variance. It does not show that every kind of exposure estimation error produces attenuation. Systematic error, error related to outcome predictors, or a nonlinear response model can produce other forms of bias.
Worked Example: Correcting a Linear Slope by Regression Calibration
Suppose exposure is measured in mg·h/L and the outcome is a clinical score in points. In a hypothetical calibration data set, the variance of latent exposure is estimated as \(100\;(\text{mg·h/L})^2\), and the variance of independent exposure error as \(25\;(\text{mg·h/L})^2\). A naïve regression of score on estimated exposure gives a slope of \(0.096\) points per mg·h/L. Under the classical-error assumptions, calculate the reliability ratio and the regression-calibration slope.
Reliability ratio. Substitute the variance estimates:
The conditional mean of latent exposure given the estimate is, under a joint normal model, \(E[X_i\mid W_i]=\mu_X+\lambda(W_i-\mu_X)\). Thus the calibrated exposure changes by \(0.800\) mg·h/L for each 1 mg·h/L change in the estimated exposure. The naïve slope is \(\lambda\beta\), so:
Under the stated model, the corrected slope is \(0.120\) points per mg·h/L: a 1 mg·h/L difference in latent exposure corresponds to a 0.120-point difference in the mean score. This is an associational interpretation, not evidence that increasing exposure causes the score to change. The correction is only as credible as the variance estimates and error assumptions used to obtain the reliability ratio.
Regression Calibration
Regression calibration replaces the error-prone exposure with its conditional expectation given the observed information. With \(W_i\) alone, that predictor is \(E[X_i\mid W_i]\). In applications, it may instead be estimated from a validation subset, replicate exposure measurements, or a PK model that supplies an individual exposure distribution. The calibration model must represent how the available exposure information relates to latent exposure.
For the linear outcome model, calibration has a particularly direct justification. By iterated expectation:
Consequently, if the outcome mean is linear in \(X_i\), and the conditional mean \(E[X_i\mid W_i]\) is correctly specified, regression on the calibrated exposure preserves the slope \(\beta\) in the population. With normally distributed \(X_i\) and \(U_i\), the conditional mean is linear in \(W_i\), as used in the example. Outside that setting, the conditional mean may need a more flexible calibration model.
For nonlinear outcome models, simply substituting \(E[X_i\mid W_i]\) into the outcome function is generally an approximation. In general, \(E[g(X_i)\mid W_i]\) is not equal to \(g(E[X_i\mid W_i])\). The discrepancy can matter when the exposure-response curve is strongly curved, exposure uncertainty is large, or the outcome model uses a nonlinear link. A correction that works well for a linear mean should not automatically be assumed adequate for a nonlinear model.
Likelihood-Based Methods
Likelihood-based methods represent the exposure error and outcome model together, then estimate the outcome parameters while accounting for the unobserved \(X_i\). For example, suppose the exposure measurement model is \(f(W_i\mid X_i)\), the outcome model is \(f(Y_i\mid X_i;\theta)\), and the latent exposure distribution is \(f(X_i;\psi)\). A subject's contribution to the observed-data likelihood is:
The likelihood for independent individuals is the product of these contributions. The integral averages over exposure values compatible with the observed information; it does not substitute a single estimated exposure as if it were known. Estimation may use numerical integration, quadrature, or other computational methods appropriate to the model.
The exposure measurement model and latent exposure distribution need support. For instance, information from validation measurements or well-characterized PK data may be necessary to estimate the size of exposure error. If the exposure distribution and error variance are both unknown and the data contain only one noisy exposure measure per person, the outcome and exposure models may not identify all parameters without additional assumptions or information.
Worked Example: A Likelihood Contribution With Uncertain Exposure
Consider a hypothetical individual with estimated exposure \(W=15\) mg·h/L and score \(Y=1.80\) points. Assume \(X\sim N(10,100)\), \(W\mid X\sim N(X,25)\), and \(Y\mid X\sim N(1+0.12(X-10),4)\). The second argument of each normal distribution is its variance. Calculate this individual's outcome density conditional on the exposure estimate, at the stated parameter values.
Normal conditioning gives \(X\mid W=15\sim N(14,20)\): the conditional mean is \(10+0.8(15-10)=14\), and the conditional variance is \(100\times25/(100+25)=20\). Integrating the outcome model over this conditional exposure distribution gives a normal outcome distribution. Its mean is \(1+0.12(14-10)=1.48\) points, and its variance is \(4+0.12^2(20)=4.288\) points squared. Thus:
This is one likelihood contribution, not a fitted population slope. In a data set, the corresponding contributions would be combined across individuals, and the likelihood would be maximized over unknown parameters. The calculation illustrates the key operation: uncertainty in \(X\mid W\) contributes to the outcome distribution instead of being discarded.
Joint PK-Outcome Modeling
A joint model can estimate exposure and clinical response in one probability model. For example, let \(\eta_i\) represent individual PK parameters, let \(C_{ij}\) be concentration data at sampling time \(j\), and let \(X_i=g(\eta_i)\) be the exposure metric derived from those parameters. A joint likelihood can include the concentration model, the PK-parameter distribution, and the outcome model:
This approach carries uncertainty from concentration observations and individual PK parameters into the exposure-response analysis. Compared with using a single individual exposure estimate, it can retain information about how precise each individual's exposure estimate is. It also requires a defensible PK model, a suitable outcome model, and a likelihood that represents their data-generating structure.
A joint model is not automatically preferable. If the PK model is misspecified, or the outcome model feeds information back into individual exposure estimates in an unintended way, the resulting estimates may be misleading. It is important to check the intended role of outcome data in estimating exposure and to examine sensitivity to plausible PK and exposure-response specifications.
Worked Example: Propagating PK Exposure Uncertainty to an Outcome
For a hypothetical individual, the PK analysis gives an exposure distribution \(X\mid\text{concentration data}\sim N(8,4)\), where exposure is in mg·h/L and the variance is in \((\text{mg·h/L})^2\). Suppose the outcome model is \(Y\mid X\sim N(0.5+0.10X,1)\), with outcome measured in points. Calculate the outcome distribution after accounting for exposure uncertainty, and compare it with a plug-in calculation using \(X=8\) as exact.
The mean outcome is \(0.5+0.10(8)=1.30\) points. Integrating over the exposure distribution adds \(0.10^2(4)=0.04\) points squared to the outcome variance. Therefore:
For an observed score of 1.50 points, the integrated density is:
Treating \(X=8\) as exact gives the same mean but outcome variance \(1.00\), omitting the \(0.04\) points squared contributed by exposure uncertainty. This simple calculation shows how uncertainty can affect outcome likelihood and inference even when the central exposure estimate is unchanged.
Choosing an Approach and Checking Its Assumptions
The methods differ in what they require and what they estimate. Regression calibration is often simpler when a credible conditional mean exposure can be estimated. Likelihood-based methods represent the exposure and outcome distributions explicitly. Joint PK-outcome models can use concentration data to estimate exposure uncertainty as part of the analysis. These categories can overlap: a joint PK-outcome model is also likelihood-based when fitted by likelihood, but it uses the PK data model directly rather than relying only on a precomputed exposure proxy.
| Approach | How uncertainty is represented | Main requirement or limitation |
|---|---|---|
| Naïve regression | Uses \(W_i\) as if it were \(X_i\) | Generally biased when exposure error is material |
| Regression calibration | Uses an estimated \(E[X_i\mid\text{exposure information}]\) | Depends on a suitable calibration model; substitution is approximate for nonlinear outcome models |
| Likelihood-based correction | Integrates over latent exposure in the observed-data likelihood | Requires credible exposure, outcome, and latent-exposure distributions, with enough information to estimate them |
| Joint PK-outcome modeling | Uses concentration and outcome data with shared latent PK parameters | Requires an adequate PK model and careful specification of how outcome data inform exposure |
Useful checks include comparing individual exposure uncertainty across the exposure range, checking whether residual patterns or predictive performance vary with estimated exposure precision, and testing sensitivity to the exposure-error variance and latent exposure distribution. If an external or validation data source is available, compare predicted and observed exposure estimates there. For PK-derived exposure, inspect whether sparse concentration sampling, shrinkage, or influential sampling times produce systematically different uncertainty for some individuals.
A common failure is to assume the classical error model simply because it gives a convenient correction. Population PK estimates can have shrinkage and unequal precision; error can depend on covariates, dose history, or sampling design. A one-number reliability ratio may then be inappropriate. Diagnostics should assess the error structure that the method needs, not just the fit of the clinical outcome model.
No correction can compensate for a poorly defined exposure metric, a misspecified exposure-response relationship, or confounding in the clinical data. The exposure quantity must correspond to the outcome model's intended predictor, and the error model must reflect how the estimate was obtained. The choice among methods is therefore driven by available data and model credibility, not by a universal ranking.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Under classical additive measurement error, why is the naïve linear slope attenuated? State the reliability ratio.
- A naïve slope is 0.15 points per mg·h/L and the estimated reliability ratio is 0.75. What is the regression-calibration slope under the linear classical-error model?
- Why can substitution of \(E[X\mid W]\) be approximate for a nonlinear outcome model?
- What additional information does a joint PK-outcome model use compared with a model that treats a precomputed exposure estimate as the only exposure input?
- Name two reasons a single classical-error reliability ratio may not describe PK-derived individual exposure estimates.
Solutions. (1) The naïve slope is \(\operatorname{Cov}(W,Y)/\operatorname{Var}(W)=\beta\operatorname{Var}(X)/[\operatorname{Var}(X)+\operatorname{Var}(U)]\). The reliability ratio is \(\lambda=\operatorname{Var}(X)/[\operatorname{Var}(X)+\operatorname{Var}(U)]\), which is below one when error variance is positive. (2) \(\hat{\beta}_{RC}=0.15/0.75=0.200\) points per mg·h/L. (3) In general, \(E[g(X)\mid W]\ne g(E[X\mid W])\), so plugging in the conditional mean can fail to represent the conditional outcome mean. (4) It can use the concentration observations and PK model to represent uncertainty in individual PK parameters and the exposure derived from them. (5) Examples include unequal precision across individuals, shrinkage, and error related to covariates or sampling design.