home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!stanford.edu!bcm!convex!darwin.sura.net!paladin.american.edu!auvm!NIOBBS1.EM.CDC.GOV!SIS1
- Return-Receipt-To: <sis1@NIOBBS1.EM.CDC.GOV>
- Encoding: 51 TEXT
- X-Mailer: Microsoft Mail V3.0 (beta-2)
- Message-ID: <2B034AF1@router.em.cdc.gov>
- Newsgroups: bit.listserv.stat-l
- Date: Fri, 13 Nov 1992 10:27:00 EST
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- From: sis1@NIOBBS1.EM.CDC.GOV
- Subject: I CHECK but I don't TEST for assumptions
- Lines: 45
-
- Ed Johnson writes
-
- > I would like to know how regularly people on this list test for assumptions.
- > For example, when a t test assumes a normal distribution and equal
- > variances, how often do you run a Lilliefor test for normalcy and an F test
- > for the dif-ference of means? I'm told that it should aways be done, but
- > these tests don't seem to be a standard part of statistical packages such
- > as SAS, nor do I recall having seen these tests being mentioned in the
- > methodology of experimental re- search. Comments, please.
-
- Funny you should ask; I'm preparing a talk for the local SAS Users Group
- with the title "Using SAS to check your statistical assumptions".
-
- Although there are certainly times that I've forgotten or I've been too
- busy, I try check assumptions religiously. But I use descriptive procedures
- like residual and normal probability plots instead of formal tests. Here's
- why.
-
- A test, especially a test for normality like Lilliefors, has low power for
- small sample sizes when you need normality most, and high power for large
- sample sizes where you need normality the least (because of the Central Limit
- Theorem). A descriptive procedure will tell you how much your data deviates
- from normality and in what direction. Certain types of non-normality, like
- light tails are less serious than others; Lilliefors doesn't give you any
- clue about this.
-
- A similar comment holds about testing equality of variances. I use the rule
- of thumb that if the standard deviations or the ranges differ by a factor of
- three or more, there is cause for concern. I've seen this rule in several
- different places; if you want a reference, I can dig one out.
-
- Most of the Statistics text books I am familiar with place a strong emphasis
- on checking assumptions. Three examples are:
-
- Draper and Smith. Applied Regression Analysis;
- Montgomery. Design of Experiments; and
- Neter, Wasserman, and Kutner. Applied Linear Statistical Models.
-
-
- Steve Simon
- sis1@niobbs1.em.cdc.gov
- hhscdc!sis1/o=cdc/ou=cinc/ou=niobbs1@mhs.attmail.com
-
- Views expressed here are my own. Mention of a product or
- company name does not constitute an endorsement by NIOSH.
-