Thao-Nguyen Pham 1,2, Anna Largajolli 1, Maria Luisa Sardu 1,3, John Maringwa 1,3, Matthew Zierhut 1, Amy Cheung 1
1 Certara (Radnor, USA), 2 Normandie Universite (Caen, France), 3 affiliation at the time of the analysis (, )
Introduction: Model-based meta-analysis (MBMA) is widely used to integrate aggregate data (AD) from randomized controlled trials (RCTs) within a pharmacologically informed modeling framework to characterize dose–response relationships and comparative treatment effects. Although MBMA is traditionally performed using published AD, access to individual patient data (IPD) may offer advantages for evaluating covariate effects and reducing uncertainty. However, IPD are often unavailable, and the magnitude of benefit from incorporating IPD into MBMA remains unclear. This study aimed to quantify the added value of IPD in MBMA, particularly in the context of predictive covariate modeling, using tofacitinib in rheumatoid arthritis (RA) as an illustrative case study.
Methods: Eleven placebo-controlled RCTs evaluating tofacitinib in RA were selected from the CODEX RA clinical outcomes database [1]. The endpoint was mean change from baseline in Disease Activity Score (DAS) at week 12. A published Emax dose–response model [2] was used to describe the relationship between tofacitinib dose and DAS change. Two modeling scenarios were explored: (1) a base model without predictive covariates and (2) a model including race (percentage of Asian patients) as a predictive covariate influencing drug effect.
Because actual IPD were not accessible, a model-based simulation approach was implemented. First, the Emax model was fitted to the reported AD to obtain reference parameter estimates. IPD datasets were then simulated for each study arm assuming normally distributed outcomes while preserving study-specific characteristics (dose levels, sample sizes, and proportion of Asian patients). For the covariate scenario, individuals were stratified as Asian or non-Asian according to reported proportions, and stratified AD were derived accordingly.
A three-step framework was used to combine IPD and AD within a unified MBMA implemented in R (nlme package) [3]. Step 1 involved fitting the Emax model to IPD alone to estimate residual error variance at the individual level. Step 2 consisted of fitting the model to AD alone with residual variances fixed to reported standard errors, ensuring appropriate weighting at the study-arm level. Step 3 integrated IPD and AD into a single model while preserving their respective variance structures. All possible permutations of multiple IPD-to-AD study ratios were evaluated. For the covariate scenario, different proportions of covariate-stratified AD were assessed. Model performance was evaluated using parameter estimation ratio (relative to reference values – PER) and relative standard error (RSE).
Results: In the base model without predictive covariates, inclusion of IPD did not affect parameter estimates or precision. Across all IPD:AD ratios, estimates of Emax and ED50 remained close to reference values derived from AD alone and consistent across permutations (PER > 90%), and RSEs were comparable regardless of the proportion of IPD included (< 20%). These findings indicate that, in the absence of predictive covariates, IPD and AD provide similar information content when analyzed at the study-arm level.
In contrast, when race was included as a predictive covariate, increasing the proportion of covariate-stratified information substantially improved estimation of the race effect parameter (Erace). Parameter estimation ratios approached reference values and RSEs decreased as more stratified data were incorporated (eg, Erace parameter PER moved from 88% to almost 100% and RSE moved from 16% to 11%). Improvements in Emax and ED50 precision were more modest but consistent. Importantly, the gain in performance was driven by access to stratified covariate information rather than by IPD alone. When only non-stratified AD were available, the limited range of aggregate covariate values constrained reliable estimation of predictive effects.
The integrated modeling approach preserved appropriate weighting between IPD and AD, with residual variance components matching those obtained from separate analyses. Implementation using open-source software demonstrated feasibility without requiring complex hierarchical or likelihood-approximation methods.
Conclusions: This study shows that incorporating IPD into MBMA does not inherently improve parameter estimation when predictive covariates are absent. However, when treatment effects are modified by covariates, access to stratified covariate information, whether derived from IPD or published subgroup results, substantially enhances estimation accuracy and precision. The primary added value therefore lies in covariate stratification rather than in individual-level data per se. Encouraging publication of stratified trial results may strengthen MBMA-based inference and support model-informed drug development decisions when full IPD access is limited.
References:
[1] “Certara Rheumatoid Arthritis Clinical Outcomes Database version 20221215,” https://codex.certara.com/codex/atopic-dermatitis
[2] M. Lamba, M. M. Hutmacher, D. E. Furst, et al., “Model-Informed Development and Registration of a Once-Daily Regimen of Extended Release Tofacitinib,” Clinical Pharmacology and Therapeutics 101 (2017): 745–753.
[3] Pinheiro J, Bates D, DebRoy S, Sarkar D, R Core Team. nlme: Linear and Nonlinear Mixed Effects Models. R package version 3.1-155 2022
Reference: PAGE 34 (2026) Abstr 12300 [www.page-meeting.org/?abstract=12300]
Poster: Methodology - New Modelling Approaches