home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!paladin.american.edu!auvm!VAX.OX.AC.UK!HALPIN
- Via: UK.AC.OX.MR; 16 OCT 92 14:35:10 BST
- Message-ID: <009622E8.5B1ECA60.30065@vax.ox.ac.uk>
- Newsgroups: bit.listserv.stat-l
- Date: Fri, 16 Oct 1992 14:33:06 +0100
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- From: Brendan Halpin <halpin@VAX.OX.AC.UK>
- Subject: Re: Akaike's Info. Crit. & PROC MIXED
- Lines: 45
-
- Clark Gaylord described Akaike's Information Criterion thus:
-
- > For the "bigger is better" version of AIC:
- > -2*ln(like) -- the bigger the better, right? But, in general gets
- > large (inappropriately from Akaike's viewpoint) as p
- > increases.
- > 2*p -- this is our "penalty" term, to take away some of the
- > "natural growth" that -2*ln(like) has for large p.
- > Hence:
- > AIC = -2*ln(like) - 2*p
- >
- > I will not comment much on the appropriateness of this measure, though I
- > think it is fair to say that the jury is still out regarding its usefulness.
-
- This sounds to me a lot like Adrian Raftery's BIC statistic.
- Quoting Yu Xie, "for large samples: BIC = Lê2 - (df)logN, where Lê2
- is the log-likelihood ratio statistic, df is the associated degrees
- of freedom, and N is the sample size". This is often used --
- recently, at any rate -- in log-linear and log-multiplicative
- models. However, the "jury is out" on this one too, with (in the
- field I know anything about) people like John Goldthorpe and Robert
- Erikson claiming it leads to excessively parsimonious models,
- countered by Raftery, Mike Hout, Xie and others who argue for its
- validity.
-
- The reason I'm posting this is (a) to ask for comments on the
- relationship between AIC and BIC (it's outside my mathematical
- competence to do it myself) and (b) to solicit opinions on the
- validity and usefulness of BIC itself.
-
-
- Refs:
- Adrian Raftery (1986) `Choosing models for cross-classifications
- (Comments on Grusky and Hauser)', Am. Soc. Rev. 51.
- Michael Hout (1989) _Following in Father's Footsteps_, Harvard UP
- Yu Xie (1992) `The log-multiplicative layer effect model for
- comparing mobility tables', Am. Soc. Rev. 57.
-
-
- Brendan
-
- Brendan Halpin Email: HALPIN@VAX.OXFORD.AC.UK
- Nuffield College Phone: +44 865 278642
- Oxford OX1 1NF
- United Kingdom
-