I do not know how to do it in SPSS, but the setup of the contrasts is
straightforward and can be used in any software. For subjects of group
1 the difference w1-w2 is an estimate of  treatment difference C-T. So
these differences score +1 on the contrastvector C-T; for group 3 the
differences w1-w2 score -1 for C-T, the other groups will get score 0
on C-T. Equivalently, w1-w2 will get scores +1 an -1 for C-M in groups
2 and 5 respectively, and for T-M in groups 4 and 6. For the analysis
you need only 2 of these contrast vectors. The contrast vector for
sequence effect on C-T has +1 and -1 for group 1 and 3 respectively,
and 0 for the other groups; the sequence effects for C-M and T-M can
be modeled equivalently. No group factor should be included in the
analysis of the differences (in a properly designed experiment there
should be no differences between groups except those induced by the
treatments; but even if there were group effects, these would be
eliminated by taking differences).

Jos Jansen


"Kasper Hornbaek" <[EMAIL PROTECTED]> schreef in bericht
news:[EMAIL PROTECTED]
> Hi,
>
> A while ago I received the suggestions below from this newsgroup,
> which was most helpful. I proceeded to do an intrablock analysis,
> which worked out fine. However, upon rereading the suggestions, I am
> slightly puzzled by some of what was proposed, can anyone clarify.
>
> Jos Jansen wrote (see below for the full question and answer) "The
> intrablock analysis can be performed by regression of the
differences
> using 2 contrast vectors for treatments differences (a third would
be
> redundant), and 3 contrast vectors for sequence effects (one for
each
> treatment combination)."
>
> I understand how to calculate differences (:->), but I am not sure
> what "regression of the differences using 2 contrast vectors for
> treatments differences" means. If I were to that in, say SPSS, how
> would I proceed? I gather I should still use the grp factor (perhaps
> encoded into some dummy variable?), but how could I set up the
> contrasts that tests for treatment differences?
>
> If the above question is too basic, perhaps someone can point me to
> information that can help answer it?
>
> Thanks in advance,
>     Kasper Hornb�k
>
> ---0---
> Fra:Jos Jansen ([EMAIL PROTECTED])
> Emne:Re: Analysis of design where subjects only use a selection of
> treatment
> View: Complete Thread (3 artikler)
> Original Format
> Nyhedsgruppe:sci.stat.edu, comp.soft-sys.stat.spss Dato:2004-02-21
> 04:06:37 PST
>
>
> "Kasper Hornbaek" <[EMAIL PROTECTED]> schreef in bericht
> news:[EMAIL PROTECTED]
> > Hi,
> >
> > I have a question on how to analyse an exprimental design.
> >
> > We had three treatments (C, T, M) of which only two could be
> > administred to each subject. The subjects receieved those
treatments
> > in two weeks following each other (W1, W2). I total there were six
> > experimental groups, each defined by a treatment combination, i.e.
> >
> > grp w1 w2
> > 1 C T
> > 2 C M
> > 3 T C
> > 4 T M
> > 5 M C
> > 6 M T
> >
> > We are interested in comparing the effects of treatments, e.g. C
vs M.
> > Right now I am considering a repeated measures analysis, with week
and
> > grp as factors. However, I cannot figure out how to compose the
> > contrasts that would answer our questions.
> >
> > Any hints? Is there some other, conceptually easier approach I
could
> > take? Any litterature on this?
>
> This experimental design is a balanced incomplete block design, with
> subjects being the blocks . A description of the analysis can be
found
> in any standard book on experimental design and the analysis can
> probably best be performed by standard software for this type of
> design. There are two possibilities: intrablock analysis, and
analysis
> with 'recovery of interblock information'. The latter is, in fact, a
> mixed model analysis.
>
> For this special case one could equivalently proceed according to
the
> following lines (I suppose that treatment combinations have been
> randomly assigned to subjects). Calculate differences and sums per
> subject. The intrablock analysis can be performed by regression of
the
> differences using 2 contrast vectors for treatments differences (a
> third would be redundant), and 3 contrast vectors for sequence
effects
> (one for each treatment combination). The interblock analysis can be
> performed similarly by regression analysis of the sums. The
estimates
> from intra- and interblock analysis may eventually be combined,
using
> the inverse of estmated variances as weights; this may be not
> worthwile, since the interblock estimates probably are much less
> precise than the intrablock estimates.
>
> This analysis covers all aspects that have been touched in posts by
> Jim Clark and Donald Burrill.
>
> Jos Jansen

.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to