Selecting an Exposure Metric

Four questions to settle before an exposure-response analysis

PMX guide

Which exposure metric to carry into an exposure-response analysis: dose or PK, empirical or model-predicted, calculated at what time, and how many metrics to try.

Published

September 6, 2026

An exposure-response (E-R) analysis needs an exposure metric, and there are many to choose from:

The four questions below decide among them. Work through them in order: the answer to each one narrows what the next one is choosing between.

Out of scope. This page chooses the metric. It does not cover how to run the E-R analysis afterwards, or how to build the PK model that some of these metrics require.

A useful companion is a comprehensive regulatory and industry review of modeling and simulation practices in oncology clinical drug development (Ruiz-Garcia et al., Journal of Pharmacokinetics and Pharmacodynamics 50:147–172, 2023).

1. Dose or a PK-Based Metric?

  1. Is PK data available? If PK is unavailable for most relevant patients, then dose should be used, or a model simulation would be needed. See Section 2.

  2. Do PK concentrations add relevant information over and above dose, or vice versa?

    • If there is little overlap in exposure between different cohorts, dose may be sufficient, as PK does not add much information.
    • If doses give PK concentrations that saturate the target, the dose intensity, or the time of drug concentration above a threshold, may be more informative than metrics derived from PK.
    • If there is just one dose, a dose-related exposure metric cannot be used.
    • If one wants to integrate data across dose regimens, or predict the response to a new dose regimen, an E-R analysis based on a PK metric may be helpful.

  3. Are there confounders that would make an E-R analysis hard to interpret? If one is not confident that the concentration-response relationship is causal, or if one cannot adjust appropriately for confounding factors, then the assigned dose, or a related metric such as assigned average dose for intermittent regimens, should be considered. Particular examples of confounding where randomized dose may be better:

    • Long-term dose adjustments due to safety or biomarker data. In long-term survival studies in oncology, patients followed for a long time often also have dose reductions due to adverse events. A dose reduction driven by an adverse response can then show up as a correlation between lower exposure and better survival, which is not causal. A longitudinal PKPD model-based E-R analysis may help here, for example a time-dependent Cox regression.
    • Sicker patients may have different clearance than healthier patients, as with PD-1 inhibitors. A dose metric would be more appropriate.
    • Different pharmacology across studies or indications may warrant separate E-R analyses for each study.

2. An Empirical or a Model-Predicted PK Metric?

  1. Is the PK data available for a sufficient number of patients at the necessary time points?

    • If PK is only available for a subset of patients, a Phase 1 study but not a Phase 3, a model may be needed to predict the PK metric for all patients.
    • If PK is not sampled at the appropriate time for all patients, a PK model may be needed to predict the PK at the time point of interest.
    • If there have been dose changes and PK is not available at the relevant time during those changes, a model may be needed.

  2. How accurate is the PK model, and how reliable is the PK sampling?

    • Developing a PK model that predicts Cmax well may be challenging for orally administered drugs. Deriving the metric directly from the PK measurement requires no model assumptions.
    • The measured Cmax may also not reflect the true Cmax, if there are too few samples, if the sampling times were not well selected, or if the PK is highly variable. Where the model is deemed reliable, a model-derived metric may be preferable to an empirical one.

  3. Can the event of interest happen over a long duration of time? Adverse events, progression-free survival (PFS) and overall survival (OS) all can. If yes, model-based PK prediction may be valuable for the time-dependent E-R analysis in Section 3.

3. At What Time, and at How Many Times?

  1. Can the event of interest happen over a long duration of time, or are there intercurrent events such as dose changes that affect the PK during the time a response event occurs?

    If yes to either, a time-dependent exposure metric is recommended. Dumortier et al., 2015 is the pattern to copy: the tacrolimus concentration is calculated every day, and the organ rejection events are drawn on top of it in black. Most of the black points fall below the median concentration, and that is what says lower exposure makes an event more likely.

    Examples:

    • In that study, tacrolimus dose decreases over time, but only in the patients that do not initially have organ rejection.
    • In oncology, long-term adverse events may be more likely in patients who have efficacy and are followed for a long duration. Patients who do well may then show lower steady-state exposures, or lower exposures at the time of event, than patients who progress in the first cycle, because they were followed long enough to have a dose reduction. Lower exposure is not causing the better response.

  2. By what mechanism do we expect response to relate to exposure?

    • For chronic, long-term effects, average exposure may be most appropriate. Safety: liver toxicity that develops from prolonged exposure. Efficacy: tumour reduction under a chemotherapy.
    • For acute effects, Cmax or the concentration at the time of event may be most relevant. Safety: adverse events such as GI toxicity, or cytokine release syndrome (CRS) in response to a T cell engager, where an acute metric suits events occurring in the next few days. Efficacy: where response is rapid, as with anti-CRS therapy.

  3. Can response affect exposure?

    If response can affect exposure, for instance where sicker patients are both less likely to respond and have faster clearance, then clearance may be more predictive of response than exposure is. Clearance can then be time-dependent, and a cycle 1 exposure metric may be preferable, because cumulative or steady-state metrics will be confounded.

    Examples:

    • PD-1 inhibitors, where patients who respond then also reduce their clearance. Dose or a cycle 1 exposure metric is the more appropriate choice.
    • Drugs with target mediated drug disposition (TMDD), where high target levels can lead to very fast clearance of the drug.

4. How Many Metrics?

Pre-specify as much as possible. Linking several exposure metrics to the endpoint and then selecting the one that shows the most pronounced E-R relationship is data-driven metric selection, and it is not recommended: the chance of a spurious result is high.

  • What is recommended is a sensitivity analysis using other metrics, to see whether the conclusion changes.
  • Alternatively, pre-specify a couple of analyses, Cmax and Cavg say, and perform both.
Back to top