1. What Is Model-Based Endpoint Selection?
Model-based endpoint selection is the use of quantitative models to determine which measurements, time points, summaries, or clinical outcomes are most informative for answering a development question.
Traditional endpoint selection may begin with clinical convention: choose a commonly used biomarker, symptom score, laboratory measure, event endpoint, or composite outcome and then design the trial around it. A model-based approach instead asks a more fundamental question: which endpoint contains the information needed to distinguish the treatment effects that matter scientifically?
The model does not replace clinical judgment. Rather, it provides a quantitative framework for connecting treatment, exposure, biological response, disease progression, and eventual clinical outcomes.
A model-based strategy connects treatment and exposure to biological response and ultimately to the endpoint that will be measured in the trial.
2. Why Use a Model to Select an Endpoint?
A treatment can affect several biological processes simultaneously. Some measurements respond rapidly, some respond slowly, and some may be downstream consequences of an earlier biological effect. A model can make these relationships explicit.
| Challenge | What a model can contribute |
|---|---|
| Multiple candidate biomarkers | Quantify which measurements are most closely related to exposure or treatment effect |
| Delayed treatment effects | Represent time delays between exposure and response |
| Longitudinal measurements | Use the entire response trajectory rather than relying on a single time point |
| Multiple doses | Relate dose to exposure and exposure to response |
| Surrogate endpoints | Evaluate whether an intermediate measure captures information relevant to a clinical outcome |
| Time-to-event outcomes | Model how treatment changes event risk over time |
| Different candidate estimands | Clarify which treatment effect the endpoint is intended to quantify |
The benefit is not simply that a model produces a numerical answer. The model forces the development team to state the assumed relationship between treatment, exposure, biological response, and the clinical outcome.
3. What Is an Endpoint?
An endpoint is a prespecified outcome used to evaluate the effect of an intervention. Depending on the disease area and development stage, endpoints can include continuous measurements, categorical responses, counts, repeated measurements, biomarkers, clinical events, or time-to-event outcomes.
| Endpoint type | Example | Typical modeling framework |
|---|---|---|
| Continuous | Change in biomarker or clinical score | Linear or nonlinear longitudinal model |
| Binary | Responder / non-responder | Logistic or generalized linear model |
| Count | Number of events during follow-up | Poisson or negative-binomial model |
| Longitudinal | Repeated biomarker measurements | Mixed-effects or joint model |
| Time-to-event | Time to progression or clinical event | Cox, parametric survival, or other event-time model |
| Composite | Time to first occurrence of multiple events | Composite or competing-event framework |
Model-based endpoint selection is therefore broader than choosing between two biomarkers. It can involve choosing the measurement itself, the analysis time point, the longitudinal summary, the event definition, or the estimand associated with the endpoint.
4. Start With the Scientific Question
Endpoint selection should begin with the decision the study is intended to support. Different decisions can require different endpoints even when they concern the same disease.
| Development question | Potentially informative endpoint |
|---|---|
| Does the treatment affect a biological pathway? | Mechanistically proximal biomarker |
| Does the treatment produce a sustained biological response? | Longitudinal response or model-derived exposure-response measure |
| Does the treatment delay disease progression? | Progression-related time-to-event endpoint |
| Does treatment improve patient experience? | Patient-reported or functional outcome |
| Does treatment reduce clinically important events? | Clinical event or time-to-event endpoint |
| Which dose should be taken forward? | Exposure-response endpoint linked to benefit and, where appropriate, risk |
5. Think of Endpoint Selection as a Causal Chain
A useful conceptual framework is to consider how an intervention moves through several levels before producing an observable clinical outcome.
PK models describe the relationship between dose and exposure. PD models describe the relationship between exposure and response. Disease-progression or time-to-event models can then describe how changes in biological response relate to longer-term outcomes.
This creates a hierarchy of potential endpoints. A proximal biomarker may respond quickly and with relatively low variability, while a clinical endpoint may be more meaningful but require a larger sample size or longer follow-up.
6. Using PK/PD Models to Identify Sensitive Endpoints
Suppose several candidate biomarkers are available. A PK/PD model can quantify how each candidate responds to drug exposure.
A simple direct-effect model might be written as:
Here, \(E_0\) is baseline response, \(E_{\max}\) is the maximum model-predicted effect, and \(EC_{50}\) is the concentration associated with half of the maximum effect.
If two candidate biomarkers have very different exposure-response relationships, the model can help distinguish them quantitatively.
| Candidate | Model behavior | Potential interpretation |
|---|---|---|
| Biomarker A | Strong exposure-response relationship with a broad dynamic range | May provide useful information about pharmacologic activity |
| Biomarker B | Weak exposure-response relationship | May be less informative about exposure-driven treatment effects |
| Biomarker C | Rapid response followed by plateau | May be useful for early biological activity but less informative at higher exposure |
The model does not determine clinical relevance by itself. It provides quantitative evidence about sensitivity to exposure, which can then be considered together with clinical meaning, measurement reliability, and other endpoint characteristics.
7. When the Endpoint Responds Slowly
A common reason to use models for endpoint selection is that the biological response may not track plasma concentration instantaneously.
For example, a simple indirect-response model can be written as:
Drug exposure may alter either the production rate \(k_{\text{in}}\) or loss rate \(k_{\text{out}}\), creating a delayed response.
A hysteresis relationship may therefore appear when response is plotted directly against plasma concentration. A model can distinguish an apparent lack of direct relationship from a genuine delay between exposure and effect.
8. Choosing a Longitudinal Endpoint
Many clinical measurements are repeatedly collected over time. Examples include biomarkers, symptom scores, lung function, tumor measurements, and functional assessments.
Instead of reducing the data to a single observation, a longitudinal model can describe the entire trajectory.
where \(Y_{ij}\) is the observation for subject \(i\) at time \(j\), \(f(\cdot)\) describes the expected trajectory, \(\theta_i\) represents subject-specific parameters, and \(\epsilon_{ij}\) represents residual variability.
This framework allows candidate endpoints such as the following to be compared:
- Change from baseline at a prespecified time.
- Maximum change from baseline.
- Time to a predefined response.
- Area under the response-time curve.
- Slope of change over time.
- Model-derived treatment effect on the trajectory.
The appropriate choice depends on what aspect of the disease process is most relevant to the development question.
9. Biomarkers and Surrogate Endpoints
A biomarker can be attractive as an endpoint because it may change earlier or with less variability than a clinical outcome. However, a biomarker is not automatically a valid surrogate for a clinical outcome.
A useful conceptual structure is:
The important question is not simply whether treatment changes the biomarker. It is whether information about treatment effects on the biomarker provides useful information about the clinical outcome.
| Question | What should be examined? |
|---|---|
| Does treatment affect the biomarker? | Treatment-to-biomarker relationship |
| Does the biomarker relate to the clinical outcome? | Biomarker-to-outcome relationship |
| Does treatment influence the clinical outcome through the biomarker? | Relationship between treatment, biomarker, and outcome |
| Can the biomarker support decisions about treatment effects? | Predictive and validation evidence across relevant settings |
10. Model-Based Selection of Time-to-Event Endpoints
For diseases in which clinically important events occur over time, an endpoint may be defined as the time until an event occurs rather than as a measurement at a fixed visit.
A survival model can describe the hazard function:
where \(h_0(t)\) is the baseline hazard and \(X\) represents treatment or other covariates. The parameter \(\beta\) determines how the covariate is associated with the hazard under the model.
Model-based endpoint selection can help investigate whether candidate events occur frequently enough, whether follow-up is sufficiently long, and whether different event definitions capture distinct aspects of disease progression.
| Potential endpoint | Potential advantage | Potential challenge |
|---|---|---|
| Time to first event | Uses event timing and is often straightforward to define | May discard information from subsequent events |
| Time to disease progression | Can capture disease-control effects | Requires a clearly defined progression rule |
| Overall survival | Directly clinically meaningful | May require long follow-up |
| Composite time-to-event | Can increase event frequency | Components may differ substantially in clinical importance |
11. Endpoint Selection and the Estimand
Endpoint selection should be consistent with the estimand: the precise treatment effect the study intends to estimate.
For example, a treatment effect might concern the difference in mean change at a specified time, the probability of response, or a treatment effect on time to an event.
| Endpoint formulation | Possible estimand |
|---|---|
| Continuous change from baseline | Difference in mean change between treatment groups |
| Binary response | Difference or ratio of response probabilities |
| Repeated outcome | Treatment effect on the longitudinal trajectory |
| Time-to-event | Treatment effect on event-time distribution or hazard |
| Composite endpoint | Treatment effect on time to the first qualifying component |
This matters because changing the endpoint can change the scientific question. A model may show that two endpoints are highly correlated, but they may still correspond to different estimands and different clinical interpretations.
12. Sensitivity, Variability, and Information
One practical reason to use models for endpoint selection is to understand how much information an endpoint contains about the treatment effect.
For a simple two-group comparison of a continuous endpoint, the approximate standardized treatment effect is:
where \(\Delta\) is the treatment-group difference and \(\sigma\) represents the relevant standard deviation.
An endpoint can therefore become more statistically informative when it has a larger expected treatment difference, lower residual variability, or both.
13. Measurement Error Matters
An endpoint can be biologically relevant but difficult to measure precisely. Measurement error can reduce the ability to detect treatment effects.
A simple measurement model is:
where \(\epsilon\) represents measurement error.
If the measurement error is large relative to the biological treatment effect, the observed endpoint may be substantially less informative than the underlying biological variable.
Model-based endpoint selection should therefore consider:
- Within-subject measurement variability.
- Between-subject variability.
- Assay or instrument precision.
- Inter-rater variability where applicable.
- Missing measurements.
- Timing of measurements.
- Changes in measurement procedures over development.
14. Simulation as an Endpoint-Selection Tool
One of the most powerful applications of model-based endpoint selection is simulation. A candidate model can be used to generate hypothetical trial data under different endpoint definitions.
For example, suppose three candidate endpoints are under consideration. A simulation can evaluate how each endpoint behaves under plausible treatment effects, variability, dropout, sampling schedules, and disease trajectories.
Simulation can compare candidate endpoints under the same assumptions about treatment effects, variability, missing data, sampling, and trial design.
The purpose is not to manufacture a favorable result. The purpose is to understand how reasonable assumptions about the disease and treatment translate into operating characteristics for competing endpoint definitions.
15. Endpoint Selection and Sample Size
Endpoint choice directly affects trial efficiency. For a simple comparison of two means, a rough sample-size relationship is:
where \(n\) represents sample size, \(\sigma\) represents variability, and \(\Delta\) represents the treatment difference of interest.
This relationship illustrates why endpoint characteristics matter. A noisy endpoint may require many more subjects than a precise endpoint for the same target effect.
However, sample size should not be the only criterion. A clinically meaningful endpoint may appropriately require more subjects or longer follow-up than a surrogate or proximal biomarker.
16. Worked Example: Selecting Among Three Candidate Endpoints
Consider a hypothetical drug development program with three candidate endpoints: Endpoint A is a proximal biomarker, Endpoint B is a functional measure, and Endpoint C is a clinical event.
Suppose a longitudinal exposure-response model provides the following hypothetical results.
| Characteristic | Endpoint A | Endpoint B | Endpoint C |
|---|---|---|---|
| Response to exposure | Strong | Moderate | Indirect |
| Time to response | Short | Moderate | Long |
| Measurement variability | Low | Moderate | Event variability |
| Clinical interpretation | Biological activity | Functional benefit | Clinical outcome |
| Expected event frequency | High | High | Lower |
Step 1: Define the development question
Suppose the immediate question is whether the selected dose produces meaningful biological activity. Endpoint A may be highly informative for this question because it responds strongly to exposure.
Step 2: Examine downstream relationships
Suppose the model also indicates that Endpoint A is associated with Endpoint B, while the relationship between Endpoint B and the eventual clinical outcome remains uncertain. This distinguishes biological activity from evidence of clinical benefit.
Step 3: Consider the clinical decision
If the decision concerns dose selection, an exposure-response endpoint may be especially useful. If the decision concerns clinical benefit, the downstream functional or clinical endpoint may be more directly relevant.
Step 4: Simulate the candidate trial
The three candidate endpoints can then be evaluated under plausible treatment effects, variability, dropout, follow-up, and sample sizes.
The function above is conceptual rather than a universal scoring formula. Its purpose is to emphasize that endpoint selection is multidimensional rather than determined by a single statistical property.
17. Joint Models and Multiple Endpoint Streams
Sometimes the most informative approach is not to select one endpoint in isolation. A joint model can connect multiple outcome processes.
For example, a longitudinal biomarker can be linked to a time-to-event outcome:
where \(m_i(t)\) represents the subject's underlying longitudinal biomarker trajectory. The parameter \(\alpha\) describes the model relationship between the latent biomarker level and event hazard.
Such models can help determine whether an early longitudinal measurement contains information about a later clinical event.
18. Model Selection Is Part of Endpoint Selection
The endpoint cannot be separated completely from the model used to analyze it. A continuous endpoint analyzed at one time point can answer a different question from the same measurement analyzed as a longitudinal trajectory.
| Endpoint representation | Modeling choice | Information captured |
|---|---|---|
| Single time point | ANCOVA or related model | Treatment difference at a specific time |
| Repeated measurements | Mixed-effects model | Trajectory over time |
| Time to threshold | Survival model | Timing of achieving a defined response |
| Count over follow-up | Poisson / negative-binomial model | Frequency of events |
| Multiple linked outcomes | Joint model | Relationships among longitudinal and event processes |
Therefore, endpoint selection should specify not only what is measured but also how the resulting data will represent the treatment effect.
19. Missing Data and Endpoint Choice
An endpoint that requires frequent assessments may be more vulnerable to missing data than an endpoint requiring only occasional measurements.
This is especially important for longitudinal endpoints. If the probability of missing a measurement is related to disease status, treatment response, tolerability, or other variables, the observed trajectory may differ systematically from the underlying trajectory.
Model-based planning should therefore consider:
- Expected dropout rates.
- Timing of discontinuations.
- Intermittent missing measurements.
- Potential informative dropout.
- Intercurrent events.
- How the chosen analysis handles incomplete trajectories.
Simulation is particularly useful here because the consequences of different missing-data mechanisms can be explored before the trial begins.
20. How Should a Model-Based Endpoint Be Evaluated?
A candidate endpoint should be evaluated against the purpose for which it is being considered. A useful evaluation framework includes several dimensions.
| Dimension | Key question |
|---|---|
| Biological relevance | Does the endpoint reflect a process affected by the intervention? |
| Clinical relevance | Does the endpoint relate meaningfully to how patients feel, function, or survive? |
| Exposure sensitivity | Does the endpoint respond to drug exposure in a quantifiable manner? |
| Precision | How much measurement and residual variability does it contain? |
| Timing | Does it change within the time frame needed for the development decision? |
| Feasibility | Can it be collected reliably in the intended population and setting? |
| Robustness | How sensitive are conclusions to plausible model assumptions? |
| Interpretability | Can the resulting treatment effect be communicated clearly? |
21. A Practical Model-Based Endpoint Selection Workflow
- Define the development decision. Determine what the trial must establish or inform.
- Define the estimand. Specify the treatment effect of interest, population, endpoint variable, and relevant intercurrent-event strategy.
- List plausible endpoints. Include biomarkers, functional outcomes, clinical events, longitudinal measures, and composites where scientifically appropriate.
- Develop a quantitative model. Use PK, PD, longitudinal, disease-progression, survival, or joint modeling as appropriate.
- Characterize treatment sensitivity. Estimate how candidate endpoints respond to treatment or exposure.
- Characterize variability. Quantify between-subject, within-subject, residual, and measurement variability.
- Evaluate timing. Determine whether the endpoint responds rapidly enough and whether the proposed assessment schedule captures the relevant response.
- Consider clinical meaning. Determine how the endpoint relates to patient benefit, disease progression, or another decision-relevant outcome.
- Simulate the trial. Compare candidate endpoints under realistic assumptions about treatment effects, sampling, dropout, and other sources of uncertainty.
- Perform sensitivity analyses. Evaluate how endpoint performance changes under alternative model assumptions.
- Finalize the endpoint and analysis strategy. Document the scientific rationale, endpoint definition, timing, estimand, and analysis model.
22. Limitations of Model-Based Endpoint Selection
Quantitative modeling can improve endpoint selection, but it does not eliminate uncertainty. Several limitations should be considered.
- Model assumptions matter. Different structural assumptions can produce different predictions.
- Historical data may not represent the future trial. Population, treatment background, disease severity, and measurement procedures may change.
- Correlation does not establish surrogate validity. An endpoint can correlate strongly with a clinical outcome without reliably capturing the treatment effect on that outcome.
- Model uncertainty should be distinguished from parameter uncertainty. A narrow confidence interval around parameters does not guarantee that the model structure is correct.
- Extrapolation can be uncertain. Predictions outside the range of observed exposure, response, or follow-up may depend strongly on assumptions.
- Clinical relevance cannot be reduced to a statistical criterion. The most statistically sensitive measurement is not necessarily the most meaningful endpoint.
23. A Simple Framework for Comparing Candidate Endpoints
A practical endpoint-selection framework can be organized around five questions:
| Question | What to examine |
|---|---|
| 1. Does it respond? | Magnitude and shape of treatment or exposure-response relationship |
| 2. Is the response measurable? | Precision, assay performance, residual variability, and missingness |
| 3. Is the timing appropriate? | Onset, duration, delay, and sampling schedule |
| 4. Is it clinically meaningful? | Relationship to patient benefit, disease progression, or clinical outcomes |
| 5. Can it support the decision? | Ability to estimate the intended treatment effect with acceptable uncertainty |
This framework is intentionally broader than a ranking system. Different development questions can lead to different endpoint choices, and a single endpoint may be useful for one decision but inappropriate for another.
24. Key Takeaways
- Model-based endpoint selection uses quantitative models to connect treatment, exposure, biological response, disease progression, and clinical outcomes.
- The process should begin with the scientific and development decision rather than with a preferred statistical test.
- PK/PD models can identify endpoints that are sensitive to drug exposure and can also help determine when those endpoints should be measured.
- Longitudinal models can use the full response trajectory rather than reducing a repeated measurement process to a single time point.
- Time-to-event models can help evaluate event frequency, timing, and alternative definitions of clinically important outcomes.
- A biomarker that predicts clinical outcome is not automatically a valid surrogate for the treatment effect on that outcome.
- Endpoint selection should be aligned with the estimand and the treatment effect the study is intended to estimate.
- Measurement error, missing data, timing, and between-subject variability can substantially affect endpoint information content.
- Simulation allows candidate endpoints to be compared under realistic assumptions about treatment effects, variability, dropout, sampling, and follow-up.
- Statistical sensitivity is only one component of endpoint selection. Biological relevance, clinical meaning, feasibility, and interpretability also matter.
- Model uncertainty should be explicitly considered rather than treating model-based predictions as directly observed facts.
- The most useful endpoint is the one that provides information appropriate to the scientific decision the trial is designed to support.
Where to Go Next
A natural progression is to study exposure-response modeling, followed by longitudinal PK/PD models, disease-progression models, biomarker qualification, surrogate endpoint evaluation, joint longitudinal-survival models, and clinical trial simulation.
The next tutorial can build directly on these concepts by showing how exposure-response modeling can be used to select a dose and identify a clinically informative response endpoint, including model development, simulation, and trial-design implications.