home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!stanford.edu!bcm!convex!darwin.sura.net!jvnc.net!gmd.de!Germany.EU.net!mcsun!news.funet.fi!hydra!klaava!tarkkone
- From: tarkkone@klaava.Helsinki.FI (Lauri Tarkkonen)
- Newsgroups: bit.listserv.stat-l
- Subject: Re: Outliers
- Keywords: Parameter estimation, outliers
- Message-ID: <1992Nov11.222750.5428@klaava.Helsinki.FI>
- Date: 11 Nov 92 22:27:50 GMT
- References: <Nov.11.16.46.06.1992.15885@gandalf.rutgers.edu>
- Organization: University of Helsinki
- Lines: 25
-
- In <Nov.11.16.46.06.1992.15885@gandalf.rutgers.edu> kluger@gandalf.rutgers.edu (Dr. Avi Kluger) writes:
-
- >I was wondering if any of the readers know of material regarding
- >the actual (as opposed to hypothetical) impact of outliers on
- >estimation of correlations. There is one article by Orr et al
- >in Personnel Psychology 1991 vol 44 no 3 which reports that various
- >outlier deletion methods had only a small impact on average rho
- >estimation and various effects on the variance of the estimates.
- >Do you know of any other recent emprical treatment of the question?
-
- >Avi Kluger
- >Institute of Management and Labor Relations
- >Rutgers University
- >(908) 932 5823
-
- Have you heard about influence curves. This means you study the change
- in correlation if a boservatin is removed from the data:
-
- Say z(x,y) = r(x,y) - r
- r = correlation for the whole data, r(x,y) is the correlation of the
- data minus that observation. The function r(x,y) can be written by
- means of n, x, y, r, mx, yx, sx, sy. (m for mean and s for standard
- deviation.).
-
- - Lauri Tarkkonen
-