home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: bit.listserv.stat-l
- Path: sparky!uunet!seas.smu.edu!convex!news.oc.com!spssig.spss.com!nichols
- From: nichols@spss.com (David Nichols)
- Subject: Multiple comparisons in SPSS
- Message-ID: <Jul21.230637.48089@spss.com>
- Date: Tue, 21 Jul 1992 23:03:19 CUT
- Nntp-Posting-Host: spssrs2.spss.com
- Organization: SPSS Inc.
- Lines: 51
-
-
- Several posts have been made in the not too distant past about using
- multiple comparison, range or post-hoc tests (depending on the preferred
- test(s) and vocabulary). In one of these there was a reference to not
- being able to do these for repeated measures factors on current software.
- This is not entirely true with regard to SPSS.
-
- The MANOVA procedure on all platforms other than DOS (SPSS/PC+) has a
- subcommand called CINTERVAL. This subcommand will allow you to produce
- simultaneous confidence intervals for each set of parameters, be they
- between or within subjects (within subjects is what MANOVA calls
- repeated measures factors). The alpha level can be specified, and the
- confidence intervals can be either Scheffe or Bonferroni type intervals.
- Multivariate intervals are also available, should they be desired.
-
- Given the way MANOVA is structured, you cannot simply ask it for all
- pairwise comparisons and get Bonferroni or Scheffe intervals; in each
- design that it runs, MANOVA estimates as many parameters for each
- effect as there exist degrees of freedom. Thus in order to estimate
- all pairwise comparisons, multiple runs with different contrasts are
- required. This is not difficult, and the Scheffe intervals can be used
- straightaway, since they involve simply the number of degrees of
- freedom in an effect and an F-distribution. The Bonferroni intervals
- are a bit more complicated, for two reasons (one of which will go away
- as of release 5). The first reason is that the divisor for alpha for
- the Bonferroni intervals is the number of degrees of freedom for the
- effect, not the number of pairwise comparisons. Thus, if one wants to
- do all pairwise comparisons with a Bonferroni correction, the alpha
- specified to the program must be adjusted to make it give what is
- desired. For example, if there are four groups, there are 6 different
- pairwise comparisons. In this case there are three df, so the divisor
- for alpha would be 3. Thus in order to get 95% Bonferroni intervals
- for all pairwise comparisons, alpha would really have to be set at
- .025 instead of .05. The other problem (which exists in current
- systems, but has already been fixed in the portable code and will be
- correct for version 5) is that there is a bug in the Bonferroni
- intervals such that they cut the alpha in half (you must ask for
- alpha of twice what you want in order to get the correct results).
-
- For those interested in doing these tests on repeated measures, there
- is another wrinkle in MANOVA: it will not allow nonorthogonal contrasts
- for WSFACTORS. However, this can easily be worked around. I have a
- description of how to do this online and will happily send it to
- anyone who would like it.
-
-
- --
- David Nichols Statistical Support Specialist SPSS, Inc.
- Phone: (312) 329-3684 Internet: nichols@spss.com Fax: (312) 329-3657
- *******************************************************************************
- Any correlation between my views and those of SPSS is strictly due to chance.
-