Gustaf Wellhagen (1,2), Bengt Hamrén (1), Maria Kjellsson (2), Magnus Åstrand (1)
(1) Quantitative Clinical Pharmacology, Early Clinical Development, IMED Biotech Unit, AstraZeneca, Gothenburg, Sweden, (2) Department of Pharmaceutical Biosciences, Uppsala University, Sweden
Background: Urinary albumin-to-creatinine ratio (UACR) is a common biomarker for drugs in the renal space as an alternative to estimated glomerular filtration rate (eGFR) because it typically has a faster response that allows for shorter studies. However, UACR is highly variable both between individuals and between visits within an individual. Results from a placebo-adjusted change-from-baseline analysis could be biased by variability in the placebo arm. Mixed-effect Model Repeated Measures (MMRM) can be an approach to model such variable data by making fewer/no assumptions regarding the placebo response. Instead, at each visit a new placebo and drug response is estimated independent of other visits. By assuming a dose-response relationship, information can be carried over between dose arms to improve predictions.
Objectives: To investigate the precision and accuracy of placebo-adjusted change from baseline of UACR with different methods: MMRM with or without a dose-response relationship.
Methods: A true dose-response model following an Emax shape with varying ED50 (2, 4, 8, 16, 32, 64, 128 mg) was assumed. Different time-courses of the drug effects were investigated (direct, linear, exponential). For each case, a number (n=1000) of 16-week studies (samples at week -2, -1, 0, 2, 4, 6, 8, 10, 12, 14, 15 and 16) were simulated with placebo and three dose arms (0, 10, 30 and 100 mg). A first-order autoregressive model (AR1) was assumed for the correlation of residuals. The simulated values of UACR were log-transformed changes from baseline. The sample size was titrated to a power of 95% for detecting a 40% reduction in log(UACR) between the highest dose arm and placebo at end-of-study.
In the traditional MMRM analysis, each visit and dose arm had a separate estimate of both the placebo response and the drug effects.
In the MMRM with dose-response analysis, each visit had a separate estimate of the placebo response and Emax, but a shared ED50 parameter, thereby saving two parameters per visit but adding one global.
The precision was assessed through the size of the estimated standard errors of the placebo-adjusted change from baseline, and accuracy through the size of the bias in the same endpoint. Both simulations and estimations were performed in R version 3.2.4 [1]. The nlme() package was used to fit the MMRM with dose-response relationship.
Results: The MMRM with dose-response had lower standard errors for the estimates of placebo-adjusted change from baseline, especially for the lower doses at higher true values of ED50 (mean standard errors at the end of study reduced by 15-60%, 15-25% and 5-10% for the 10, 30 and 100 mg dose arms respectively). The traditional MMRM was unbiased while the MMRM with dose-response had a slight bias at the lower doses, the bias relative to the true effect of the highest dose amounted to up to 2.5% for the two lower doses. The bias increased with higher true ED50.
Conclusion: MMRM with an incorporated dose-response relationship offers an improvement in precision over traditional MMRM analyses. The improvement is mostly seen for lower doses.
References:
[1] R Core Team (2016). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/.
Reference: PAGE 28 (2019) Abstr 9152 [www.page-meeting.org/?abstract=9152]
Poster: Methodology - New Modelling Approaches