E. Niclas Jonsson 1, Siv Jönsson 1, Emma Hansson 1, Joakim Nyberg 1
1 Pharmetheus Ab (Uppsala, Sweden)
Objectives
Covariates in pharmacometric models have two important objectives[1]. First, they support mechanistic understanding by helping to identify the physiological drivers of drug disposition and effect. Second, they support clinical decision making such as dose selection and labeling. These two objectives target different quantities of interest. Mechanistic understanding often requires isolating the impact of specific physiological drivers, while clinical guidance typically benefits from evaluating the overall effect a covariate has.
In current practice, the distinction between these objectives is often blurred. This is partly due to a tradition of prioritizing mechanistic parsimony and partly due to a lack of standardized terminology to support the distinction of the two types of covariate effects. Consequently, the mathematical method (the estimator) is frequently selected before the scientific question (the estimand) is clearly defined. We propose that the ICH E9(R1)[2] estimand framework provides a structured vocabulary to reconcile these two perspectives, ensuring that covariate effect models are fit-for-purpose.
Methods
Applying the ICH E9(R1) framework, we define two distinct covariate estimands:
The conditional estimand: This estimand targets the isolated effect of a covariate, i.e. the sensitivity of a parameter to the covariate, keeping all other covariates constant. It is the natural choice for establishing or verifying mechanistic hypotheses and is typically estimated via multivariable models obtained from data-driven selection methods such as the SCM[3].
The unconditional estimand: This estimand targets the total impact of the covariate, i.e. the overall impact of the covariate regardless of any other covariates. It mimics the real-world clinical setting where other covariates vary according to their natural correlations. This is the relevant target for drug labeling and is most directly estimated via univariable models or derived from multivariable Full Random Effects Models (FREM)[4].
To demonstrate the implications of these estimands, a theoretical framework was established to describe the difference between conditional and unconditional covariate effects in the presence of covariate correlation. This was supported by simulations utilizing a cohort from the National Health and Nutrition Examination Survey (NHANES)[5] (n=7276), focusing on the relationship between body mass index (BMI) and body surface area (BSA). We evaluated the risk of misinterpretation when conditional covariate coefficients are used as proxies for (the overall) unconditional clinical impact for dose recommendations in the labelling.
Results
The analysis confirms that unless the correlation between covariates is zero, conditional and unconditional effects are not equivalent. The discrepancy is a mathematical inevitability where the unconditional effect captures both the isolated effect and the shared influences of correlated variables.
Theoretical and simulated results show that even at modest correlations (e.g., ⍴=±0.25), conditional coefficients can deviate by 25%–30% from the unconditional coefficients.
In the BMI/BSA case study (⍴=0.74), a model using conditional coefficients in isolation incorrectly suggested that a single dose level would be appropriate across all BMI quintiles. In contrast, the unconditional estimand revealed that over 50% of subjects in the lowest and highest quintiles would be significantly over- or under-dosed if a flat dose were applied.
While a complete multivariable model remains accurate for population simulations, individual conditional coefficients are non-robust and highly sensitive to the correlation structure of the covariates included in the model.
Conclusions
With the common practice of reporting conditional coefficients in isolation, e.g. within abstracts, reports, publications, drug labels, or forest plots, there is a risk that mechanistic signals are misinterpreted as total clinical impacts. This can result in misleading conclusions regarding the clinical relevance of a covariate.
We suggest that a more transparent approach is to report both the conditional and unconditional effects. This evolution in practice aligns with the ICH M15[6] focus on the “question of interest” within Model-Informed Drug Development (MIDD). By defining the estimand (objective) first, and the estimator (method) second, the pharmacometrician ensures that the quantitative evidence generated is optimized to improve patient care and regulatory communication.
References:
[1] Jonsson, E. Niclas; Nyberg, Joakim; Mentré, France. Objective First, Method Second: Why the Estimand Definition comes First in Pharmacometric Covariate Modeling. (Submitted).
[2] International Council for Harmonisation. CH E9(R1) addendum on estimands and sensitivity analysis in clinical trials to the guideline on statistical principles for clinical trials. E9(R1), Geneva, Switzerland: International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use, 20 November 2019.
[3] Jonsson EN, Karlsson MO. Automated Covariate Model Building within NONMEM. Pharm Res 1998; 15: 1463–1468.
[4] Yngman G, Bjugård Nyberg H, Nyberg J, et al. An introduction to the full random effects model. CPT: Pharmacometrics & Systems Pharmacology; 11. Epub ahead of print 2022. DOI: https://doi.org/10.1002/psp4.12741.
[5] Centers for Disease Control and Prevention (CDC). National Center for Health Statistics (NCHS). National Health and Nutrition Examination Survey 2017-March 2020 Prepandemic Data Files., https://wwwn.cdc.gov/nchs/nhanes/search/datapage.aspx?Component=Demographics&Cycle=2017-2020 (2020).
[6] International Council for Harmonisation. ICH M15 guideline on general principles for model-informed drug development – Final version. M15 (Step 4), Geneva, Switzerland: International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use, 29 January 2026.
Reference: PAGE 34 (2026) Abstr 11875 [www.page-meeting.org/?abstract=11875]
Poster: Methodology - New Modelling Approaches