1. What Is a Binary PK/PD Response?
A binary response has only two possible outcome categories. Examples include responder versus nonresponder, seizure versus no seizure, remission versus no remission, or occurrence versus nonoccurrence of a predefined event during a specified observation period.
In PK/PD modeling, the central question is often not simply whether drug concentration changes, but whether increasing exposure changes the probability of a particular response.
For a binary endpoint, the PD model describes how exposure changes the probability of one of the two possible outcomes.
2. What Makes a Response Binary?
Suppose the observed outcome for subject \(i\) is \(Y_i\). A binary endpoint can be represented as:
The corresponding response probability is:
Unlike a continuous endpoint, the model is therefore concerned with the probability of observing \(Y=1\). Each individual contributes a 0 or 1, while the model describes how the probability of a 1 changes with exposure and other predictors.
| Binary endpoint | Possible coding | Example interpretation |
|---|---|---|
| Clinical response | 1 = responder, 0 = nonresponder | Achieved a predefined response criterion |
| Remission | 1 = remission, 0 = no remission | Met the study-defined remission criterion |
| Adverse event | 1 = event, 0 = no event | Experienced the specified event during follow-up |
| Biomarker threshold | 1 = above threshold, 0 = below threshold | Exceeded a prespecified biomarker level |
| Responder threshold | 1 = clinically meaningful improvement, 0 = otherwise | Reached a predefined change from baseline |
The scientific definition of the endpoint is important. A statistical model cannot compensate for a binary endpoint whose definition, timing, or clinical interpretation is unclear.
3. What Does PK Contribute?
The PK component describes how administered dose produces concentrations or exposure over time. The PD component then uses an appropriate exposure measure to explain variation in response probability.
A general PK/PD structure can be written as:
Here, \(C(t)\) represents concentration over time and \(X\) represents an exposure metric used by the PD model. Depending on the scientific question, \(X\) might be a concentration at a particular time, average concentration, \(C_{\max}\), AUC, trough concentration, or another model-derived exposure measure.
For a time-dependent response, the model may instead use the predicted concentration \(C(t)\) directly, allowing the probability of response to vary over time.
4. The Logistic PK/PD Model
The most common probability model for a binary endpoint is logistic regression. It models the log odds of response as a function of exposure.
Equivalently, the response probability can be written as:
This transformation is useful because the linear predictor \(\alpha+\beta X\) can take any real value, while the inverse-logit transformation guarantees that the resulting probability lies between 0 and 1.
Interpreting the parameters
- \(\alpha\) is the intercept and determines the baseline level of response probability when \(X=0\).
- \(\beta\) describes how exposure changes the log odds of response.
- A positive \(\beta\) means that increasing exposure increases the modeled odds of response.
- A negative \(\beta\) means that increasing exposure decreases the modeled odds of response.
The odds ratio associated with a one-unit increase in exposure is:
The odds ratio describes a multiplicative change in odds, not a constant change in probability. Consequently, a given change in exposure can produce different absolute changes in probability depending on the starting probability.
5. Modeling a Dose- or Exposure-Dependent Probability
A simple logistic model assumes that exposure has a linear effect on the log-odds scale. This can be useful, but many pharmacologic exposure-response relationships are nonlinear.
One way to introduce a characteristic exposure level is to use an \(E_{\max}\)-type function inside the logit:
In this formulation, the exposure-response relationship is nonlinear on the probability scale and on the log-odds scale because the exposure effect approaches a limiting value.
An alternative formulation can place an \(E_{\max}\)-type function directly on the probability scale, provided the resulting parameterization guarantees valid probabilities:
Here \(p_0\) represents the baseline probability and \(p_{\max}\) represents the asymptotic probability under the specified model.
6. Logistic Versus Probit Models
A binary response does not require the logistic link. Another common choice is the probit link:
where \(\Phi^{-1}\) is the inverse standard normal cumulative distribution function.
| Model | Link function | Interpretation |
|---|---|---|
| Logistic | \(\log[p/(1-p)]\) | Models log odds of response |
| Probit | \(\Phi^{-1}(p)\) | Models response probability on a latent normal scale |
| Complementary log-log | \(\log[-\log(1-p)]\) | Can be useful for asymmetric probability relationships and event-type modeling |
Logistic and probit models often produce similar fitted probabilities over the range of observed data. Their coefficients are on different scales, however, so coefficients should not be compared directly without accounting for the link function.
7. When Response Probability Changes Over Time
Sometimes the binary response is not a single endpoint measured once. Instead, the probability of response can change as drug concentration changes over time.
A time-dependent model can be written as:
The PK model supplies \(C(t)\), and the PD model converts concentration into the probability of response at time \(t\).
For example, if drug concentrations rise after dosing and subsequently decline, a positive exposure-response relationship can produce a corresponding rise and fall in the predicted probability of response.
When concentration varies over time, a PK/PD model can produce a time-varying probability of binary response.
This approach is conceptually different from simply relating one observed binary endpoint to an AUC value. The time-dependent formulation attempts to preserve the temporal relationship between exposure and response.
8. Adding Patient and Study Covariates
Response probability may depend on more than drug exposure. Baseline disease severity, treatment group, demographic characteristics, biomarkers, concomitant treatments, or other predictors can be incorporated into the PD model.
A logistic model with exposure and a covariate \(Z\) can be written as:
An exposure-by-covariate interaction can be added when the exposure-response relationship is expected to differ between groups:
The interaction term allows the slope associated with exposure to change as \(Z\) changes.
9. Individual Variability in Binary PK/PD Models
Patients can differ substantially in both PK and PD behavior. One patient may have greater exposure because of lower clearance, while another may have a different probability of response at the same exposure because of differences in pharmacodynamic sensitivity.
A simple individual-level logistic model can represent PD variability through a subject-specific random effect:
where \(b_i\) represents an individual-specific deviation from the typical response tendency.
In a population PK/PD analysis, variability can therefore arise from multiple sources:
- PK variability: individuals differ in clearance, volume, absorption, or other PK parameters.
- PD variability: individuals differ in their response at a given exposure.
- Residual variability: observed outcomes contain stochastic variation even after accounting for modeled predictors.
The distinction between PK and PD variability is important because a difference in response can arise either because subjects experienced different exposure or because they respond differently to the same exposure.
10. How Binary PK/PD Data Are Organized
A typical dataset might contain one or more exposure variables together with a binary outcome:
| Subject | Exposure | Covariate | Response |
|---|---|---|---|
| 001 | 2.0 | 0 | 0 |
| 002 | 3.5 | 1 | 1 |
| 003 | 5.0 | 0 | 1 |
| 004 | 7.5 | 1 | 1 |
| 005 | 10.0 | 0 | 1 |
The exposure variable could be an observed or model-derived PK metric. In a population PK/PD workflow, it is often preferable to use individual exposure estimates generated by the PK model rather than treating sparse observed concentrations as if they were error-free exposures.
When repeated binary observations are collected from the same individual, the dependence among observations must also be considered. Standard logistic regression assumes independent observations unless additional structure is introduced.
11. Worked Example: Exposure and Probability of Response
Suppose a hypothetical drug is evaluated using a binary clinical response. The PK analysis provides an exposure measure \(X\), expressed as average concentration in mg/L. A logistic PD model is specified as:
Step 1: Calculate the log odds at \(X=2\)
Step 2: Convert log odds to probability
At an exposure of 2 mg/L, the model predicts approximately a 26.9% probability of response.
Step 3: Increase exposure to \(X=4\)
Therefore:
At an exposure of 4 mg/L, the predicted probability of response is 50%.
Step 4: Increase exposure to \(X=6\)
At 6 mg/L, the predicted probability is approximately 73.1%.
| Exposure \(X\) | Log odds | Predicted probability |
|---|---|---|
| 0 mg/L | −2.0 | 11.9% |
| 2 mg/L | −1.0 | 26.9% |
| 4 mg/L | 0.0 | 50.0% |
| 6 mg/L | 1.0 | 73.1% |
| 8 mg/L | 2.0 | 88.1% |
This example illustrates why the logistic model is useful for binary PK/PD data: a linear exposure effect on the log-odds scale produces a nonlinear relationship between exposure and probability.
12. What Is the Exposure for a 50% Response Probability?
For a simple logistic model:
When \(p=0.5\), the log odds are zero because:
Therefore, the exposure producing a 50% predicted response probability satisfies:
and hence:
For the worked example, \(\alpha=-2\) and \(\beta=0.5\), giving:
Thus 4 mg/L is the modeled exposure at which the predicted response probability is 50%.
13. Why Odds Ratios Are Not Probability Differences
Suppose \(\beta=0.5\). The odds ratio associated with a one-unit increase in exposure is:
This means that the modeled odds of response are multiplied by approximately 1.65 for a one-unit increase in exposure, holding other variables constant.
It does not mean that the response probability increases by 65 percentage points.
For example, increasing a probability from 10% to 20% is very different from increasing it from 70% to 80%, even though both are 10-percentage-point increases. Logistic models naturally account for this because the relationship is defined on the log-odds scale.
14. Why Exposure Is Often More Informative Than Dose
Two patients receiving the same dose can have different concentrations because of differences in clearance, absorption, distribution, organ function, body size, drug interactions, or other factors.
A dose-response analysis therefore asks a somewhat different question from an exposure-response analysis.
| Relationship | Primary question |
|---|---|
| Dose → response | How does administered dose relate to response probability? |
| Exposure → response | How does the achieved drug exposure relate to response probability? |
| Concentration → response | How does concentration at a specified time relate to response probability? |
| Time-varying concentration → response | How does the probability of response change as concentration changes? |
PK/PD modeling provides a framework for separating the dose-to-exposure relationship from the exposure-to-response relationship.
15. Choosing a Binary PK/PD Model
The choice of model should reflect the scientific question and the amount and quality of information available in the data.
- Define the binary endpoint. Specify exactly what constitutes response and when it is assessed.
- Choose the exposure metric. Decide whether dose, concentration, AUC, \(C_{\max}\), trough concentration, or another exposure measure is scientifically appropriate.
- Explore the exposure-response relationship. Examine observed response rates across exposure ranges before imposing a particular functional form.
- Select a link function. Logistic regression is common, but probit and other links may be appropriate.
- Consider nonlinear exposure effects. Use flexible or mechanistically motivated relationships when a simple linear logit is inadequate.
- Add important covariates. Account for scientifically relevant patient or study characteristics.
- Assess model adequacy. Examine predictions, residual or diagnostic information, calibration, and parameter plausibility.
- Use the model for prediction carefully. Predictions outside the exposure range supported by the data depend strongly on the assumed model.
16. How Should a Binary PK/PD Model Be Evaluated?
A model that produces reasonable parameter estimates is not automatically an adequate model. Evaluation should consider whether predicted probabilities reproduce the observed response patterns.
- Observed versus predicted response rates: compare observed and model-predicted proportions across exposure groups or bins.
- Calibration: examine whether predicted probabilities correspond reasonably to observed frequencies.
- Discrimination: when relevant, evaluate how well the model separates subjects with and without the response.
- Parameter plausibility: check whether estimated effects and exposure-response shapes are scientifically reasonable.
- Influential observations: assess whether a small number of subjects dominate the fitted relationship.
- External or predictive evaluation: where possible, evaluate performance in data not used to estimate the model.
17. What Binary PK/PD Models Do Not Tell Us Automatically
Binary outcomes are useful but discard information compared with the underlying continuous or time-to-event process. Several limitations should therefore be considered.
- Dichotomization loses information. A continuous response converted into responder/nonresponder may discard differences between individuals near and far from the threshold.
- Endpoint definition matters. Changing the response threshold can change the apparent exposure-response relationship.
- Timing matters. A response assessed at one time point may not represent the complete pharmacodynamic trajectory.
- Exposure measurement contains uncertainty. Treating estimated exposure as known without accounting for PK uncertainty can affect inference.
- Model form matters. Different link functions or exposure-response functions can produce different extrapolations.
- Correlation matters. Repeated binary outcomes from the same individual are not generally independent.
- Predicted probabilities are model-based. They should not be interpreted as observed individual-level certainty.
18. A Practical Binary PK/PD Modeling Workflow
- Define the binary endpoint. Specify the response criterion and assessment time.
- Develop or obtain the PK model. Estimate individual or population exposure as appropriate.
- Choose the exposure measure. Base the choice on pharmacology and the expected mechanism of action.
- Explore response by exposure. Visualize observed response rates across exposure ranges.
- Fit an initial binary-response model. A logistic model is often a useful starting point.
- Investigate nonlinear exposure-response relationships. Consider Emax-type or other scientifically motivated functions when appropriate.
- Evaluate covariates. Assess clinically relevant predictors and potential exposure-by-covariate interactions.
- Assess model adequacy. Examine calibration, predictions, diagnostics, and parameter plausibility.
- Perform simulations if needed. Use the PK/PD model to explore response probabilities under alternative dosing or exposure scenarios.
- Communicate uncertainty. Distinguish observed response rates from model-based predictions and clearly state the assumptions underlying extrapolation.
19. Using Binary PK/PD Models for Simulation
Once a PK/PD model has been estimated and evaluated, it can be used to simulate response probabilities under different exposure scenarios.
For a logistic model:
Changing \(X\) produces a corresponding change in the predicted probability. If a PK model predicts exposure under a particular dosing regimen, the exposure can be passed into the PD model to obtain a predicted response probability.
This creates a useful bridge between pharmacokinetic modeling and clinical decision-making simulations, while keeping clear that the final predictions remain conditional on the estimated model.
20. Key Takeaways
- A binary PK/PD endpoint has two possible outcomes, commonly coded as 0 and 1.
- The PD model describes the probability that the response occurs rather than predicting a continuous effect.
- The PK model supplies concentration or another exposure measure to the response model.
- Logistic regression models the log odds of response and guarantees predicted probabilities between 0 and 1.
- The logistic coefficient represents a change in log odds; its exponential is an odds ratio.
- An odds ratio is not the same as a percentage-point change in response probability.
- Probit and other link functions can also be used for binary PK/PD relationships.
- Exposure-response relationships can be linear on the logit scale or nonlinear through Emax-type and other functions.
- The exposure associated with a 50% response probability can be derived directly from a simple logistic model.
- Covariates can explain differences in baseline response probability or differences in exposure-response sensitivity.
- PK and PD variability represent different sources of between-subject differences in observed response.
- Repeated binary observations require methods that account for within-subject dependence.
- Binary endpoint definitions and assessment times can substantially influence the resulting exposure-response relationship.
- Model evaluation should consider calibration, predictive performance, parameter plausibility, and scientific interpretation.
- PK/PD simulations can connect dosing regimens to predicted exposure and ultimately to predicted response probabilities.
Where to Go Next
A natural progression is to study PK/PD Models for Continuous Endpoints, followed by count-data models, time-to-event models, categorical responses, and population PK/PD approaches.
For binary responses specifically, the next level of detail is to explore nonlinear logistic exposure-response models, Emax models for probability of response, covariate effects, individual variability, repeated binary outcomes, and simulation-based evaluation of alternative dosing regimens.