Test Of Concept By Simulation: Comparing Response-Rate Findings Between Parallel And Titration Designs
Background. Between-subject variability (BSV) in dose-response for efficacy (E) and toxicity (T) leads to difference in optimal dose among patients. Therefore parallel group trials (PRL) conducted at fixed doses are likely to underestimate the true response rate (R) due to over-dosing some patients while under-dosing some others at each tested dose. In drug development, early knowledge of a clinically meaningful R helps establish a realistic expectation for the potential medicine and facilitates crucial investment decisions. A titration design (TRN) that mimics clinical practice would help achieve this. Conceivably, the difference between the R estimated from these two designs can be influenced by factors such as therapeutic window, dose selection, measurement error for both E and T, and co-variance for E and T. Expected outcome can be simulated for both designs in scenarios defined in these aspects.
Objective. The objective of this work is to compare the response rate and the robustness and efficiency of its estimation, in terms of accuracy and precision, between PRL and TRN trials.
Methods. Trial data were simulated for hypothetical drugs whose E and T can each be described by a simple Emax model. Drugs with wide or narrow therapeutic index were tested at three dose levels. For TRN, dose was increased step-wise if E is lower than 0.5 and T is lower than 0.2. Subjects with E of at least 0.5 and T lower than 0.2 (at the final dose, in the case of TRN) were considered as responders. Simulations of PRL and TRN trials were conducted for various levels of BSV (25% to 75%) in the mean doses producing 50% of maximum E (ED50) and T (TD50). Impact of correlation (up to 0.9) between ED50 and TD50 was also considered. A small measurement error was included for both E and T. For each scenario and design, replicate trials of various sample sizes were simulated to obtain median and 90% confidence interval of R. Necessary sample size for achieving a desirable precision of R estimate at 90% confidence level was compared between the two designs. The R obtained using both designs were compared to the true value.
Results and Conclusion. Preliminary findings show that response rates estimated by TRN trials are up to twice that estimated by PRL trials, in the scenarios tested here. Further, lower sample size is generally required to achieve similar level of precision. These results support the preferential use of TRN where applicable.