CMStatistics 2016: Start Registration
View Submission - CMStatistics
B0334
Title: Interpretation and evaluation of circular effects in projected normal regression models Authors:  Jolien Cremers - Utrecht University (Netherlands) [presenting]
Kees Mulder - Utrecht University (Netherlands)
Irene Klugkist - Utrecht University (Netherlands)
Abstract: Currently, coefficients for projected normal regression models come on two bivariate components. This enables predictions of a circular effect, but does not allow for the interpretation and evaluation of the size of the effect on the circle. We introduce new types of circular regression coefficients that do allow us to assess the size of this effect. These coefficients are straightforward to compute from the original bivariate output of an MCMC sampler. The bivariate coefficients allow us to distinguish between location and accuracy effects on the circle. We propose a measure to assess whether there is any effect on the circle at all. Subsequently we check whether this effect is most likely an accuracy or a location effect. The performance of these measures and of the new circular regression coefficients is investigated using a simulation study. The new measures are illustrated using example datasets for a standard multivariate regression and a mixed effects model. In addition, we use the Bayesian toolbox for model assessment and evaluation to check whether the assumption of projected normality is met, compare the fit of several models and test order constrained hypotheses for these example datasets.