Assessment of pharmacometric and statistical analysis methods in dose-finding trial design: application in diabetes drug development

Aims: To evaluate the performance of different pharmacometric and statistical methods of analysis when evaluating power of dose-finding trial designs for treatment of type 2 diabetes, with differing levels of model uncertainty (due to limited prior data in patients).

Methods: A population PK/PD model was used to simulate exposure-response relationships for HbA1c and glucose in a typical Phase 2 patient population with diabetes. Although the true underlying model is assumed to be known (an indirect response model of fasting blood glucose linked to HbA1c ), different levels of uncertainty are simulated as scenarios with varying confidence in the model parameters. The uncertainty in estimates of Emax and EC50, ranging from 0 to 200% CV, reflect the amount of prior knowledge in patients. Doses for the simulations were selected based on the current parameter estimates. 1,000 trial replicates were simulated for each combination of sample size and number of dose levels. The simulated data were then used for longitudinal and dose/exposure-response modeling by both pharmacometrics and statistics to calculate power. Power was calculated for various metrics, including ability to estimate model parameters and the target dose, using the different models. Simulations are visualized using Shiny, an R package.

Results: In general, power increased with sample size and number of dose levels, and decreased with increasing uncertainty. For example, using an exposure-response model for a 5-dose design, the sample size required for at least 80% power to estimate a target dose whose true typical HbA1c response is within 0.15% of the desired response was 39/arm for the lowest uncertainty, and 53/arm for the highest uncertainty (Figure 1). For Emax models, exposureresponse models typically provided slightly greater power than dose-response models. When using the linear model, power was significantly lower due to model misspecification. Results from other analyses are forthcoming. 

Figure 2 

Conclusion: Emax models based on exposure and dose performed comparably, although exposure-response models offer greater advantage when uncertainty is high. Simple linear regression was extremely poor, suggesting that power is low when using a model that does not well describe the underlying dose-response relationship. Further, model uncertainty has significant impact on the design of a dose-finding Phase 2 study. Thus, leveraging prior data in T2DM patients may have significant cost-saving benefits.