Shrinkage-Induced Bias in Sequential PD Fits

Why PK Shrinkage Matters for Sequential PD Fits

In a sequential analysis, PK data are used first to estimate individual PK parameters, and the resulting estimates are then used to predict each subject’s exposure for the PD analysis. When PK data are sparse or noisy, individual empirical Bayes estimates (EBEs) are pulled toward the population mean. This shrinkage can reduce the spread of predicted exposures across subjects. If the PD fit treats those compressed predictions as known, the fitted exposure-response parameters may no longer have the interpretation intended on the true exposure scale.

The effect is not a universal movement in one direction. A simple multiplicative compression of a predictor rescales a slope and potency parameter in predictable ways. Noisy exposure estimates can instead attenuate a slope. In nonlinear models, the consequences also depend on the exposure range, the model, and which parameters are estimated together. The central task is therefore to identify what kind of shrinkage or prediction error is present before interpreting the PD estimates.

A Simple Shrinkage Model

Let \(\theta_i\) be a subject’s PK parameter, such as log clearance, and suppose the population model is \(\theta_i\sim N(\mu,\omega^2)\). To isolate the mechanism, suppose the individual PK data can be summarized by an estimate \(z_i\) satisfying \(z_i\mid\theta_i\sim N(\theta_i,s^2)\), with the same known standard error \(s\) for each subject. The conditional mean of \(\theta_i\) given \(z_i\) is

The factor \(\lambda\) is between zero and one. When individual PK information is strong, \(s^2\) is small and \(\lambda\) is near one, so the estimate remains close to \(z_i\). When individual information is weak, \(s^2\) is large and \(\lambda\) is small, so the estimate is pulled toward \(\mu\). In a nonlinear PK model the posterior estimate need not be this simple, but the same qualitative pull toward the population distribution can occur.

If a predicted exposure is a function of the individual parameter, \(C_i(t)=C(t;\theta_i)\), the sequential analysis often uses \(C(t;\hat{\theta}_{i,\mathrm{EB}})\). Variation in the EBEs is usually less than variation in the underlying individual parameters, so predicted exposure variability can be compressed too. The relationship is not always a simple multiplication: for nonlinear PK models, changing a parameter can alter both the level and shape of the concentration-time profile.

How a Compressed Predictor Changes a Slope

Consider a linear PD relationship \(Y_i=a+\beta C_i+\epsilon_i\), where \(C_i\) is the exposure relevant to the response and \(\epsilon_i\) has mean zero. If the sequential fit uses \(C_i^*=cC_i\), with the same known constant \(0<c<1\) for every subject, then the same mean response can be written

Therefore the fitted slope against the compressed predictor is \(\beta^*=\beta/c\). This is a slope that is larger in magnitude when \(c<1\); it is not attenuation. The numerical coefficient changes because its predictor is expressed on a compressed scale. If the compression factor is known and the coefficient is converted back to the original exposure scale, the corresponding value is \(c\beta^*=\beta\).

This result requires a common multiplicative compression, a linear mean model, and a correctly specified intercept. It does not establish that every use of EBEs inflates slopes. In practice, compression may vary by subject or time, and EBEs can also contain estimation error. Those conditions can produce different bias patterns.

Why Noisy Exposure Estimates Can Instead Attenuate a Slope

A different mechanism arises when the predictor contains additive estimation error. Suppose \(W_i=C_i+u_i\), where \(u_i\) is independent of \(C_i\) and of the PD residual, has mean zero, and has variance \(\sigma_u^2\). Under the linear model with centered variables, the large-sample regression slope of \(Y_i\) on \(W_i\) is

The ratio multiplying \(\beta\) is less than one when \(\sigma_u^2>0\), so this classical noisy-predictor mechanism attenuates the slope toward zero. The assumptions matter: errors must be independent in the specified way, and the expression concerns an additive-error model. EBEs are not generally independent additive errors; they are calculated using a population distribution and may have subject-specific uncertainty. Thus a compressed EBE predictor cannot be analyzed by automatically applying the classical attenuation result.

There is also an important calibration distinction. Under a correctly specified normal model, the posterior mean \(E(C_i\mid\text{PK data})\) is calibrated as a conditional mean. For a linear PD mean and PK data independent of the PD residual, using that conditional mean can preserve the conditional expected response. But plugging it in and treating it as an error-free covariate still omits uncertainty and may not reproduce the intended likelihood or uncertainty estimates. In nonlinear PD models, the response at a posterior mean exposure generally differs from the posterior mean response; the earlier material on nonlinear exposure plug-in addresses that separate issue.

Consequences for Potency Estimates

Potency parameters in nonlinear PD models are also tied to the exposure scale. For an inhibitory or stimulatory Emax relationship, a common mean function is

Here \(E_0\) is the baseline response, \(E_{\max}\) is the maximum change from baseline, and \(EC_{50}\) is the exposure producing half of that maximum change. If the predictor is scaled to \(C^*=cC\), the identical response curve is represented by \(EC_{50}^*=cEC_{50}\), because \(C^*/(EC_{50}^*+C^*)=C/(EC_{50}+C)\). A fit using compressed exposure values can therefore return a correspondingly compressed apparent \(EC_{50}\). This is a change in the exposure scale, not evidence by itself that biological potency has changed.

Diagnostics, Alternatives, and Limitations

Shrinkage is a warning about the information in individual PK estimates, not a stand-alone proof that a PD estimate is biased. Assess the PK-to-exposure step before interpreting the sequential PD fit.

• Quantify parameter shrinkage. A commonly used descriptive measure for a random effect \(\eta\) is \(1-\mathrm{SD}(\hat{\eta}_{\mathrm{EB}})/\omega\), where \(\omega\) is the population standard deviation. Values nearer one indicate greater compression. This measure is approximate and can be affected by the distribution of the EBEs; it is not a direct estimate of PD bias.
• Compare exposure distributions. Inspect the spread of individual predicted exposures across subjects at relevant times, and compare it with uncertainty-aware or population-model predictions. Also examine whether compression differs across subjects or concentration ranges.
• Check the PD fit on the scale it uses. Plot residuals and predictions against the fitted exposure predictor. A narrow predictor range can make the slope or \(EC_{50}\) sensitive to small changes in exposure estimates, even when ordinary residual plots appear acceptable.
• Test sensitivity to the PK-to-PD handoff. Compare sequential estimates obtained with alternative defensible exposure summaries or with uncertainty propagated from PK. A substantial change indicates that the PD result depends on how individual PK uncertainty is handled.

The main alternatives differ in how they handle individual PK uncertainty. A plug-in sequential fit uses point estimates of PK parameters or exposure and is computationally convenient, but treats those values as fixed. An uncertainty-propagating sequential method carries information about their uncertainty into the PD analysis. A simultaneous PK/PD likelihood estimates PK and PD parameters together and can use both data sources in the joint fit; its validity still depends on suitable model specification and identifiability. These approaches are not interchangeable in every design, and none automatically corrects a misspecified PK or PD model.

The direction of distortion is not universal. Pure multiplicative compression gives an increased linear slope and a proportionally decreased \(EC_{50}\) when those coefficients are reported on the compressed scale. Additive independent predictor error attenuates a linear slope. Subject-specific shrinkage, nonlinear PK predictions, limited exposure coverage, and dependence between PK and PD subject effects can yield more complicated results. The last issue requires explicit modeling rather than a blanket correction factor.