Jones, M. C.
|DOI (Digital Object Identifier) Link:||http://dx.doi.org/10.1016/0167-9473(91)90049-8|
|Google Scholar:||Look up in Google Scholar|
It is a simple matter to correct for the well-known variance inflation property of nonnegative kernel density estimates whereby the estimated distribution's variance exceeds that of the sample. But should we bother? Asymptotic mean integrated squared error considerations, developed here for the first time, suggest we may. However, we observe that the difference variance correction makes is, in most practical instances, negligible. Even when this is not so, exploratory conclusions would rarely be affected and, on occasions when this is not so either, variance correction can have a slight tendency to obscure potentially important features of the density. An exception to all this is estimation of the normal density for which correcting for variance inflation is certainty appropriate. This author retains a personal preference for continuing with uncorrected kernel density estimates, but the main message of the paper is the relative indifference to whether or not variance correction is employed.
|Item Type:||Journal Article|
|Copyright Holders:||1991 Elsevier Science B.V.|
|Keywords:||bimodality; exploratory data analysis; mean integrated squared error; normal density; skewness|
|Academic Unit/Department:||Mathematics, Computing and Technology > Mathematics and Statistics
Mathematics, Computing and Technology
|Depositing User:||Sarah Frain|
|Date Deposited:||22 Mar 2011 11:46|
|Last Modified:||18 Jan 2016 10:06|
|Share this page:|