Artikel
A Simulation Based Approach to Estimating Sample Sizes for Clinical Trials
Suche in Medline nach
Autoren
-
Lukas Baumann
- Institute of Medical Biometry and Informatics, University of Heidelberg, Heidelberg, GermanyCenter for Medical Statistics, Informatics and Intelligent Systems, Medical University of Vienna, Vienna, AustriaNovartis Pharma AG, Basel, Switzerland
-
Martin Posch
- Center for Medical Statistics, Informatics and Intelligent Systems, Medical University of Vienna, Vienna, Austria
-
Björn Bornkamp
- Novartis Pharma AG, Basel, Switzerland
| Veröffentlicht: | 26. Februar 2021 |
|---|
Gliederung
Text
Background: Sample size determination is a vital step in planning a clinical trial. However, more and more study designs are used in practice for which no closed formula for the design's power can be derived, e.g. various adaptive designs and Bayesian designs. In these cases, repeated simulations of the trial are necessary to estimate the sample size. But as modern trial designs become more complex, their simulation also becomes computationally more expensive and hence more time-consuming. Systematic investigation of a design under different assumptions may even become infeasible.
Methods: To avoid wasting computational resources, and as a consequence time, the objective of our work was to derive an algorithm that estimates sample sizes accurately with as few simulations of a trial as possible. We show that sample size estimation via simulation can be viewed as finding the root of a stochastic function. We propose the Bayesian Local Linear Regression (BLL) algorithm for sample size estimation via simulation, which uses repeated locally linear probit regressions that are fitted to the outcomes of simulated trials at various sample sizes. The parameters of the probit regression are estimated using Bayesian methods to utilize prior knowledge about the shape of the power function. The sample size is estimated from the probit regression model.
Results: We conducted a simulation study where we compared our algorithm to other stochastic root finding algorithms. The BLL algorithm is able to estimate the sample size for various scenarios more precisely than the other algorithms when the trial is simulated 100, 250, or 1000 times in total. A further advantage of our algorithm compared to other stochastic root finding algorithms is that due to the model-based nature we can express the uncertainty of the power at the estimated sample size.
Conclusion: We conclude that the BLL algorithm is a helpful tool for researchers to save time when they have to use simulations to estimate the sample size for a clinical trial.
The authors declare that they have no competing interests.
The authors declare that an ethics committee vote is not required.
