Clémence Rigaux (1), Bernard Sébastien (1)
(1) Sanofi R&D, Paris, France
Introduction: Since 2005, pharmaceutical companies have been required by the Food Drug Agency to complete formal safety and efficacy pediatric trials for 21 new non-insulin drugs under the Pediatric Research Equity Act. However, recruitment for pediatric trials in Type II Diabetes Mellitus (T2DM) is very challenging (1), necessitating new approaches for reducing the sample sizes of pediatric trials in T2DM. In this work we assessed if a longitudinal Non-Linear-Mixed-Effect (NLME) analysis of T2DM trial results could be more powerful and so require less patients than standard analyses. These standard analyses currently used are Last Observation Carried Forward (LOCF) + analysis of variance (ANOVA) analysis, and Mixed-effects Model Repeated Measures (MMRM) analysis, which is a repeated measures analysis including treatment-by-visit interaction effect and that accounts for missing data phenomenon. The longitudinal NLME analysis is expected to be more precise and more powerful than standard analyses as it includes information from all observed time points and as the difference between drugs could be summarized by only one drug effect parameter.
Objectives: to compare the power of MMRM analysis, LOCF + ANOVA analysis at end of study, and longitudinal NLME analysis in the assessment of results of a simulated T2DM trial.
Methods: Studies in T2DM were simulated, with glycated hemoglobin (Hba1c) as main endpoint, 24 weeks duration, 2 arms, 75 patients per arm, HbA1c measurements at week 0, 8, 12 and 24 as in standard clinical trials in T2DM, and assuming a mean placebo effect of -0.1% HbA1c and a mean treatment effect of -0.5% HbA1c. Hba1c change was modelled using 3 scenarios: 2 with a structured model of negative exponential progression with steady state reached at week 12 (A) or week 20 (B), and an unstructured (ie not coming from a mathematical model) scenario (C). Inter-individual-variability was accounted for. Dropout was also modeled, assuming a Missing At Random (MAR) phenomenon and a global dropout rate at week 24 of about 10%.
1000 trials were simulated and for each trial, statistical significance of a between-group difference in HbA1c change from baseline at week 24 was assessed, using the 3 types of analyses:
- LOCF + ANOVA at week 24
- MMRM approach, then computation of treatment difference at week 24
- Longitudinal NLME modelling, using a structured model of negative exponential shape, then prediction of treatment effect at week 24.
Then the power of each type of analysis was computed as the % of the 1000 trials with significant tests, and these powers were compared. Two types of power were computed: in case of bad convergence of an analysis for a trial, the test was considered either missing or non-significant.
Results: The LOCF + ANOVA analysis slightly under-estimated the magnitude of drug effect and was less powerful than the MMRM and NLME analyses in all scenarios (mean bias between +0.010 and +0.024 %HbA1c depending on the scenario, and power always < 78%). For structured scenarios A and B, similar power was observed between MMRM and NLME analyses (around 83% for scenario A and 85% for scenario B depending on the method and type of power calculation), and for unstructured scenario C, NLME analysis was a little more powerful than MMRM (92% versus 88%) but was associated with a slight overestimation of magnitude of drug effect (mean biais -0.044 %HbA1c).
A slightly more precise estimation of the drug effect parameter was obtained with the NLME method compared to the two other methods (standard error ≤0.172 versus between 0.179 and 0.189 %HbA1c).
Further explorations showed that adding some HbA1c measurements at week 4 and 52 led to a gain in power for NLME compared to MMRM (gain ~+8% for scenarios B & C), and that the type 1 error was a little inflated for MMRM (9%) and for NLME (7 to 10%).
Conclusions: The longitudinal modelling analyses MMRM and NLME were more powerful than the LOCF + ANOVA analysis at week 24. The NLME analysis gave slightly more precise drug effect estimations than the 2 other methods. However, it tended to overestimate the magnitude of drug effect and it was more powerful than standard MMRM analysis only in some scenarios, with an increased gain in power in presence of additional time-points. These initial results suggest that NLME analyses may help to reduce the required sample sizes in T2DM pediatric studies, provided that sufficient HbA1c assessments are planned.
References:
[1] Nadeau KJ, Anderson BJ, Berg EG, Chiang JL, Chou H, Copeland KC, Hannon TS, Huang TT, Lynch JL, Powell J, Sellers E, Tamnorlane WV, Zeitler P. Youth-Onset Type 2 Diabetes Consensus Report: Current Status, Challenges, and Priorities. Diabetes Care. 2016 Sep; 39(9):1635-42.
Reference: PAGE 28 (2019) Abstr 8814 [www.page-meeting.org/?abstract=8814]
Poster: Drug/Disease Modelling - Paediatrics