Feeding Exposure-Response Models Into Trial Simulation

From an Exposure-Response Model to Patient Outcomes

An exposure-response model becomes useful for trial planning when it can produce outcomes for the kinds of patients who may enter a proposed trial. The simulation needs more than a fitted curve: it needs regimen-specific exposures, patient covariates, a rule for calculating the model-predicted response, and—when predicting realized outcomes—a rule for generating residual variation around that response.

The handoff is conditional on the model being used in the right context. The exposure metric must match the one used to fit the exposure-response relationship; the covariates must have consistent definitions and units; and the model’s outcome, time point, and population must correspond to the trial question. The simulator then applies that model to each simulated patient, rather than treating a group-level prediction as if it were every patient’s outcome.

Specify the Patient-Level Calculation

Let \(C_i\) denote the exposure metric for patient \(i\), \(X_i\) the patient’s covariates, and \(\hat{\theta}\) the estimated exposure-response parameters. A simulator first calculates a model-predicted outcome, or a parameter of the outcome distribution, from \(C_i\) and \(X_i\). It then follows the fitted model’s outcome scale and distribution to generate a realized outcome if that is the desired simulation target.

For a continuous outcome with a normal residual model, one common specification is:

Here \(m(C_i, X_i; \hat{\theta})\) is the predicted conditional mean, and \(\hat{\sigma}\) is the estimated residual standard deviation on the outcome scale. This specification assumes that the residuals follow the stated distribution and, if the model assumes constant residual variance, have the same variance across the simulated patients. Residuals should only be generated in a way consistent with the fitted model’s assumptions.

For a binary outcome, the model may instead specify a probability. For example, with a logistic model, \(\operatorname{logit}(p_i)=\eta_i\), where \(p_i=P(Y_i=1\mid C_i,X_i)\) and \(\operatorname{logit}(p)=\log(p/(1-p))\). The simulator converts \(\eta_i\) to \(p_i\), then generates a binary outcome from a Bernoulli distribution with that probability. The link scale is not the outcome scale: a residual term should not be added to \(p_i\) unless that is what the model specifies.

The exposure inputs must come from a representation appropriate to the trial question. They may be individual exposure predictions supplied by a linked pharmacokinetic component, or prespecified exposure values for a demonstration. In either case, the simulator must use the same metric and units as the exposure-response model. If the model was fitted using AUC in a particular time interval, substituting a concentration or an AUC over a different interval changes the model input and invalidates the handoff.

Why Patient-Level Simulation Matters

For a trial-level quantity such as the mean outcome, patient-level predictions must be aggregated over the simulated patients. If \(m_i=m(C_i,X_i;\hat{\theta})\), then the average predicted outcome is \(\sum_{i=1}^{n}m_i/n\). If the exposure-response relationship is nonlinear, this is generally not equal to evaluating the response function at the average exposure and covariate values. The individual exposure and covariate distribution therefore matters even when the target is an average.

A short derivation shows how residual variation enters a simulated group mean under a simple model. Suppose \(Y_i=m_i+\varepsilon_i\), with residuals having mean zero and variance \(\sigma^2\), and suppose they are independent across patients conditional on the supplied exposures and covariates. Then:

The equality follows because \(E(\varepsilon_i\mid C_i,X_i)=0\). Under the stated independence and common-variance assumptions, the conditional variance of the group mean is:

This result is conditional on the simulated exposures and covariates and on the stated residual assumptions. It does not account for uncertainty in the estimated exposure-response parameters or for other sources of variability that have not been included in the simulation.

Binary Outcomes and Trial Simulation

For a binary outcome, calculate a probability for each patient before drawing the outcome. Suppose \(\eta_i\) is the fitted linear predictor. The corresponding probability is \(p_i=1/(1+\exp(-\eta_i))\). A simulated event indicator can be generated as \(Y_i=1\) when a uniform random draw \(U_i\) on \([0,1]\) is less than \(p_i\), and \(Y_i=0\) otherwise. The expected event proportion is the average of the individual probabilities; the realized event proportion is based on the generated indicators.

Implementation Checks and Common Failure Modes

Before relying on trial-simulation output, check the path from regimen to exposure to outcome. Each link should be consistent with the model used to estimate the exposure-response relationship and with the intended trial population and endpoint.

Common warning signs include exposures outside the range supporting the fitted relationship, predictions outside a plausible outcome range, probabilities below 0 or above 1, and sharp changes caused by a covariate coding error. A simulation that gives identical outcomes to all patients may have omitted residual variability or incorrectly applied a group-level prediction. Conversely, adding residual noise when the target is only the model-predicted mean can obscure the quantity being reported.

A further pitfall is using the wrong scale for residual variation. In a normal model, \(\hat{\sigma}\) may be expressed in outcome units; in a logistic model, the link function defines how probabilities are obtained, and outcomes are generated as binary draws. Applying a normal error directly to a probability can produce impossible values and does not implement the stated logistic model. The simulator must follow the fitted model’s outcome distribution rather than impose a convenient but incompatible error rule.

Choosing the Simulation Target and Recognizing Limitations

A deterministic calculation of \(m(C_i,X_i;\hat{\theta})\) is appropriate when the target is a predicted conditional mean or probability. Stochastic outcome generation is needed when the target is a distribution of patient outcomes or a realized trial statistic. These approaches answer different questions; a stochastic simulation does not make the underlying mean prediction more valid, and a deterministic prediction does not describe the spread of individual outcomes.

A parametric residual model generates outcomes from the distribution specified by the fitted model. An alternative is to resample residuals, but that approach requires residuals that are suitable for resampling and does not automatically preserve a model’s covariate-dependent variance or outcome type. Neither approach repairs a misspecified exposure-response function or makes predictions reliable beyond the evidence supporting the model.

A plug-in simulation uses estimated parameters as fixed values. It can describe outcomes conditional on those estimates, but it does not represent uncertainty in them. It also does not establish that the regimen will cause the predicted response: that interpretation depends on the validity of the linked models and assumptions, as well as the clinical setting. Report the target, assumptions, exposure range, and whether results are conditional predictions or simulated realized outcomes.