CMStatistics 2017: Start Registration
View Submission - CMStatistics
Title: Mean integrated squared error comparison between kernel and maximum likelihood density estimates of normal mixtures Authors:  Dimitrios Bagkavos - University of Crete (Greece) [presenting]
Prakash Patil - Mississippi State University (United States)
Abstract: A methodological advance is presented in order to help practitioners to decide in selecting between parametric and nonparametric estimates when estimating mixtures of normal distributions. Through graphical tools and simulations numerical evidence is provided indicating that, as expected, the parametric approach is a more accurate choice when the number of mixture components is small. As the number of components increase, the difficulty in precisely estimating their number, together with the convergence issues that might occur in the implementation of the maximum likelihood algorithm, point to the direction of kernel based estimates as a more reliable option. Further investigation sheds light on the choice between simple and the variable bandwidth version of kernel density estimates.