Parameter Significance Test Using Mixture Models (PaSTUM) allowing Type-1 Error Control for Exposure-Response Modelling

In drug development, exposure-response models are widely used to inform decisions in dose optimization processes. Type I error (T1) due to mis-specified models can lead to critical and costly decisions. Therefore, a new approach called: "parameter significance test using mixture models (PaSTUM)" is proposed and compared against the standard approach (STA). A stochastic simulation estimation workstream was performed to test T1 rate, power, precision, and accuracy of drug-effect parameter estimates as well as predictive performance of the models. A total of 78 simulation scenarios for a hypothetical antidiabetic drug were investigated. For the T1 investigation, the arm allocation was randomly permutated and the AUC values were randomly sampled with replacement to each patient. For the power investigation, no arm permutation was performed, and the treatment arm remained unchanged while only the placebo patients got randomly sampled AUC values from the treatment patients. The relative root mean square error (rRMSE) and relative bias (rBias) were analyzed for predictive performance. Precision and accuracy for the drug-effect parameter were compared to the STA parameter estimates for each structural model with the most patients. PaSTUM outperformed STA regarding T1 rate inflation (20/78,78/78 > 6.53%) for PaSTUM and STA respectively. Power was marginally worse for PaSTUM. rRMSE and rBias for the full PaSTUM and STA models were similar in the power setting, but PaSTUM outperformed STA in the T1 setting. Precision and Accuracy of parameter estimates were similar for PaSTUM and STA.