Using Superpower R package to simulate power by Dr Aaron R. Caldwell

About the speaker

Aaron is an exercise physiologist with a PhD in Health, Sport, and Exercise Science from the University of Arkansas. He is currently as ORISE Postdoctoral Fellow at the United States Army Research Institute of Environmental Medicine where his current research projects are focused on human performance in extreme environments (heat, cold, and altitude). In addition, Aaron works as an applied statistician.  He has gained expertise in statistics through a Graduate Certificate in Statistics & Research Methods program while at the University of Arkansas, and is continuing his education with an MSc in Analytics through the Georgia Institute of Technology.

About the talk

Power analysis has become the sine qua non for justifying for sample sizes in experimental studies. For simple one or two sample comparisons, the process is fairly straightforward. When the experiments become more complex, such as when factorial designs are implemented, the tools used for power analysis are sparse and the calculations become more difficult. A simple solution to the problem of design complexity is just to simulate the study design in order to estimate power. However, simulation tends to require more technical knowledge and some ability to write code. Therefore, Superpower R package was created to streamline the simulation process and make simulation tools accessible for the average researcher. Currently, the package allows for Monte Carlo and “exact” simulations for factorial designs with both within and between subjects factors. This allows for power estimates for ANOVA, MANOVA, and estimated marginal means comparisons. In addition, this is a useful teaching tool as it can show how violating assumptions (e.g., homoskedasticity or sphericity) can affect both statistical power and type 1 error rates. In this presentation, I will demonstrate 1) why these principles are important 2) how Superpower, in both its R and Shiny formats, can be useful and 3) how non-simulation based functions can be used to justify your alpha level.