Bengt Hamrén, Susanne Johansson, Karin Nelander and Magnus Åstrand
AstraZeneca R&D, Gothenburg, Sweden
Objectives: ANOVA or analysis of covariance (ANCOVA) is commonly used as the primary analysis for clinical studies. For study designs with repeated measurements Mixed-Effect Model Repeated Measure (MMRM) is often used as either primary or secondary analysis. However, for studies including >1 dose strength, one drawback for both methods is that no information is obtained for doses not tested in the study. The objective of this work is to evaluate, both in terms of power and type 1 error, longitudinal dose-response modelling of all measured data with the aim of estimating treatment effect at end of study.
Methods: Clinical trial simulations of a placebo controlled multi dose HbA1c study with end of study HbA1c treatment effect as the primary endpoint were performed to compare ANCOVA, MMRM and longitudinal dose-response modelling (here LDRM) with respect to power and type 1 error. Repeated measurements of HbA1c were simulated using an indirect response model with a dose dependent inhibitory treatment effect on kin. Model parameters for simulations were set based on observed data from in-house studies. Analyses were performed both after all patients have completed the study, and also at an interim when about half of the patients had entered the study, not all of them having complete data. For the evaluation of power, 1000 studies were simulated with two efficacious doses of a hypothetical drug. Type one error was estimated from the simulation of 10000 studies, where none of the two doses had any effect.
For HbA1c the within individual variability is relatively low compared to the total variability. To broaden the scope, simulations with higher within individual variability were also performed.
Simulations were performed in R (version 3.2.2, the R Foundation for Statistical Computing) and analyses were performed in SAS (version 9.3, SAS Institute Inc., Cary, NC, USA).
Results: The simulations show that the LDRM performs in line with the ANCOVA and the MMRM, both with respect to power and with respect to type 1 error. The LDRM provided somewhat better power compared to ANCOVA, and also to MMRM, when the full study was analysed but more so at the interim. This difference was even more pronounced in the case of a higher within individual variability.
Conclusions: The longitudinal dose-response modelling can provide more informed decision making without losing power or inflating type 1 error.
Reference: PAGE 25 (2016) Abstr 5963 [www.page-meeting.org/?abstract=5963]
Poster: Methodology - Study Design