Worked Example From Exposure-Response to Trial Outcomes

From Exposure Predictions to Trial Outcomes

An exposure–response model does not, by itself, predict how many patients in a trial will respond or experience an adverse event. It maps an exposure to an outcome probability. To predict trial outcomes, the exposure distribution, exposure–response relationships, uncertainty in their parameters, and the random generation of patient outcomes must be connected in a coherent sequence.

This tutorial works through that sequence for a hypothetical treatment. The example uses a continuous exposure measure in milligrams per liter, binary efficacy and safety outcomes, and a placebo group for comparison. The numerical values are illustrative, not estimates from a real drug or study. The central distinction is between variability—differences among patients and random outcomes—and knowledge uncertainty about model parameters. They contribute differently to a trial prediction.

Specify the Linked Models and Their Assumptions

Suppose the treatment exposure for patient \(i\), \(C_i\), is represented by a lognormal population distribution. Define it as \(C_i=(2\ \text{mg/L})\exp(Z_i)\), where \(Z_i\sim N(\log 2,0.4^2)\). The two factors in this definition are intentional: the leading \(2\ \text{mg/L}\) is a scale factor, while \(\exp(Z_i)\) is a dimensionless random multiplier. Thus \(\log C_i\) is normal after expressing exposure relative to the chosen units.

For this worked simulation, assume patients’ exposures are independent draws from this distribution, conditional on the stated population model. In an application, the exposure distribution could instead come from an individual-level population PK model that includes covariates, dosing history, and residual variability. The relevant requirement at this handoff is that the exposure values passed to the response models have the same meaning and units as the exposures used to define those models.

Let \(Y_i=1\) denote an efficacy response and \(A_i=1\) denote an adverse event. The hypothetical efficacy model is logistic, as is the safety model:

Here \(C_i\) is in mg/L, so the constant \(2\) in each exposure fraction also has units of mg/L; the fraction \(C_i/(2+C_i)\) is dimensionless. For the central parameter values, take \(\beta_0=-1\), \(\beta_1=2\), \(\gamma_0=-3\), and \(\gamma_1=1.2\). These choices make the efficacy probability rise with exposure and the adverse-event probability rise more modestly with exposure over the range considered. They are illustrative assumptions, not universal properties of exposure–response models.

For the basic simulation, assume that, conditional on exposure and parameters, the efficacy and adverse-event outcomes are Bernoulli variables with the probabilities above. Assume conditional independence between the two outcomes for this example. This means the joint probability of a response and an adverse event at a given exposure is the product of their probabilities. If clinical knowledge or data indicate dependent outcomes, a joint model is needed instead. A placebo efficacy probability of \(0.30\) will be used for a simple randomized comparison. The example does not assign placebo an exposure-dependent response.

What the Exposure Distribution Contributes

For a lognormal variable, the mean is not its median. If \(Z\sim N(\mu,\sigma^2)\), then \(E[\exp(Z)]=\exp(\mu+\sigma^2/2)\). Applying this result to the specified exposure model gives:

The median exposure is \((2\ \text{mg/L})\exp(\log 2)=4.000\ \text{mg/L}\), lower than the mean because the lognormal distribution is right-skewed. This distinction matters when summarizing exposure, but neither the mean nor the median alone is generally enough to calculate the population-average response probability. In general, the response curve is nonlinear, so the response at the mean exposure need not equal the average response across exposures.

Translate Individual Exposures Into Outcome Probabilities

For a fixed exposure and fixed parameters, each model yields a probability between zero and one. The probability is an expectation for a patient with those model inputs; it is not an outcome that has already been observed. For a concrete calculation, consider four hypothetical treatment exposures: \(0.8\), \(1.5\), \(2.5\), and \(4.0\ \text{mg/L}\). They are illustrative patient values, not a claim that they summarize the population distribution.

Propagate Parameter Uncertainty

The central parameter values give conditional probabilities, but they do not represent uncertainty about the exposure–response relationships. When parameter estimates are uncertain, a simulation should draw parameters from their joint uncertainty distribution and use the same draw consistently within each simulated trial. The joint distribution matters: if estimates are correlated, independently drawing each parameter can produce the wrong uncertainty in the predicted outcomes.

A short calculation shows why parameter uncertainty changes an individual prediction. At \(C=2.5\ \text{mg/L}\), consider two hypothetical efficacy-parameter draws, each given equal weight for illustration. Draw A is \((\beta_0,\beta_1)=(-1.2,1.8)\); draw B is \((-0.8,2.2)\). Under draw A, the predictor is \(-1.2+1.8(0.5556)=-0.2000\), so \(p_E\approx0.4502\). Under draw B, it is \(-0.8+2.2(0.5556)=0.4222\), so \(p_E\approx0.6040\). Averaging over these two parameter draws gives:

This average is the outcome probability averaged over the two specified parameter scenarios, not a probability that either parameter scenario is correct. In a real analysis, parameter draws would come from the fitted model’s uncertainty distribution, including relevant parameter correlations. Exposure variability would also be represented by generating or integrating over exposures. The resulting predictive distribution can then reflect both kinds of uncertainty.

Generate Patient-Level Trial Outcomes

Once exposures and probabilities have been specified, outcomes can be generated using Bernoulli draws. For a binary outcome with probability \(p\), draw \(U\) uniformly from \([0,1]\) and set the outcome to one if \(U<p\), and zero otherwise. This converts a probability prediction into one possible patient-level realization. Repeating the procedure with new random draws produces other plausible realizations.

A trial simulation repeats this process many times. Within each replicate, it generates a trial’s exposures and patient outcomes, then calculates quantities of interest such as response proportions, risk differences, or adverse-event counts. If parameters are uncertain, a replicate first draws a parameter set and then uses it for the trial under that set. The distribution of results across replicates describes the range of trial outcomes implied by the model and its specified uncertainty.

Justification: Why Average Over Exposure and Parameters?

For a randomly selected treated patient, the population probability of response is not generally the response probability at the mean exposure. The law of total probability gives the marginal response probability by averaging the conditional probability over both exposure and uncertain parameters. If \(\Theta\) denotes the efficacy parameters and \(g(C\mid\Theta)\) represents the relevant exposure distribution conditional on those parameters, then:

This identity follows by conditioning first on \(C\) and \(\Theta\), and then averaging those conditional probabilities over their distributions. In general, because the logistic function is nonlinear, \(E[p_E(C_i\mid\beta)]\ne p_E(E[C_i]\mid\beta)\). Likewise, evaluating a model at a point estimate of \(\Theta\) need not equal averaging predictions over uncertain parameter draws. Simulation approximates the displayed integral by generating draws and averaging the resulting conditional probabilities or trial summaries.

Diagnostics, Alternatives, and Limitations

Several checks help identify failures at the handoff from exposure to outcomes. Inspect the exposure distribution for implausible values, unexpected skew, or disagreement with the intended dose and population. Plot or tabulate outcome probabilities across the exposure range and verify that they remain between zero and one and change in the expected direction. Confirm that the exposure units used in the response model match those in the PK output. A unit mismatch can leave the equations computable while making the predictions meaningless.

Compare simulated summaries with their conditional expectations where these are available. For fixed exposures and parameters, the simulated average outcome across many replicates should approach the average of the patient-specific probabilities. If it does not, check the outcome-generation code and random-number comparisons. Also check whether uncertainty has been propagated at the correct level: drawing a new shared parameter set independently for every patient can misrepresent trial-level uncertainty, while holding an uncertain parameter fixed across every replicate omits that uncertainty from the predictive distribution.

A plug-in calculation using central parameter values is simpler and can be useful for a clear conditional prediction, but it does not reflect parameter uncertainty. Evaluating the response model at mean exposure is also simple, but may misrepresent the population-average probability because of nonlinearity. Direct integration can estimate a population average without generating individual binary outcomes, which is useful when only expected outcomes are needed. Patient-level simulation is more flexible for trial-level distributions and joint summaries, but requires more assumptions and careful implementation.

The worked models are deliberately limited. They assume a specified exposure distribution, a fixed dose assignment, binary outcomes, and conditional independence of efficacy and safety. They do not represent dropout, adherence, time-to-event outcomes, exposure measurement error, covariate effects, or dependence between endpoints. Nor does a simulation guarantee that the models are adequate for a new population or trial design. Conclusions are conditional on the model structure, parameter uncertainty distribution, and assumptions used to generate data. Diagnostics can reveal inconsistencies, but they cannot establish that every assumption is true.

Common Pitfalls

• Dropping a scale factor. In this example \(C_i=(2\ \text{mg/L})\exp(Z_i)\). Its mean includes the leading factor: \(2\exp(\log 2+0.08)\ \text{mg/L}\approx4.333\ \text{mg/L}\).
• Reporting a probability as an observed outcome. A predicted probability of \(0.08840\) means an 8.840% conditional adverse-event probability under the stated model inputs; it does not imply that a patient has experienced a fraction of an event.
• Using the response at mean exposure as the population response. For nonlinear models, calculate or simulate the average of patient-specific probabilities over the exposure distribution instead.
• Confusing a simulated result with an expected result. A small trial can have an observed proportion far from the average probability. Report which quantity is being interpreted.
• Ignoring parameter uncertainty or drawing parameters at the wrong level. State whether a result is conditional on fixed parameters or averages over their uncertainty, and use a coherent parameter draw for each simulated trial replicate.