http://spssx-discussion.165.s1.nabble.com/GLM-repeated-measures-concern-question-tp5104022p5112700.html
Diana seems to know something about the subject at hand.
But I have to offer an obvious suggestion - I think the word needed
was "monotonic" and not "linear", in "reaction time is a linear function
of set size...." -- Surely the difference between sets=1 vs 2 is much
larger than the difference between sets= 31 vs. 32 -- which is what
is implied by "linear". My intuition says, That is not reasonable. My
experience says that going against that intuition results in non-linear
relations that introduce artifacts in the testing and complications in the
interpretations.
I would agree that "throwing out the linear assumption is ESSENTIAL" if
she were referring to the spacing of (1,8,32). That is not reasonable.
But she says, "Test whether log or raw set size is better linear
predictor" and I presume that "raw" refers to (1,8,32).
As to logs: The log-spacing of (1,8,32) can be taken from the
respective powers of 2: (0,3,5). That is prettly close to linear, so
it is not unreasonable to use it, if you do not have a program that
affords the luxury of using the slightly unequal log-spacing (0,3,5).
It is also justified by the experimenters' (presumable) intention of
providing equal spacing in their design.
(Of course, the log is not the only power transformation that might
be considered for spacing.)
--
Rich Ulrich
Date: Thu, 29 Dec 2011 14:23:09 +0000
From:
[hidden email]Subject: {!!! SPAM ???} Re: GLM repeated measures concern/question
To:
[hidden email]
Re: GLM repeated measures concern/question
This study follows on well-established work of Sternberg & later Triesman, Townsend where the set size is numeric and MATTERS
It is now very well known, & should be in psych101, that reaction time is a linear function of set size and the slope matters because changing conditions may change intercept but not slope or change both slope and intercept. There are strong theoretical implications concerned with parallel and serial processing and the location of processing limitations.
It may well be that log transformation will lead to better fit, but this is NOT a good reason to drop the numerical information in set size.
Test whether log or raw set size is better linear predictor.
Thrwoing out the linear assumption is ESSENTIAL
Best
Diana
On 28/12/2011 20:01, "Rich Ulrich" <rich-ulrich@...> wrote:
Do you really have reason to throw out the assumption of
equal spacing?
The set sizes (1,8,32) are obviously non-linear in spacing
when you consider simple counts, but the dimension that
matters is the dimension of impact on the measured outcome.
Thus, "equal intervals" for a drug dosage will sometimes be
in log-units, rather than raw units - as a simple example.
It is my experience that experienced researchers pick
"unequal intervals" in time, in setting up repeated measures,
in such a fashion to approximate equal intervals in outcome.
I've usually had "time" for the repeated measures I've tested.
And it is a mistake not to take advantage of the intelligent
design when doing the analyses.
It is my observation that nominally unequal numeric intervals,
whether they are time or design elements like yours, are
usually (similarly) constructed with the intent of producing
equal intervals of outcome/ response. So you might try the
simple analysis, assuming equal intervals, and see if fits.