On Fri, Aug 10, 2012 at 11:29 AM, Johannes Keyser <[log in to unmask]> wrote:
> Thank you very much for answering so quickly!
>
> Just to clarify: Given a design including one between-subject factor GROUP and within-factors A and B, which error term should be used for testing the single contrast column A1,B1?
Error Term 1.
> Asking more general: How can I decide which error term is appropriate? What is the key concept needed for this decision?
The key concept is what are you comparing. If you are comparing a
single condition against 0 (e.g. A1, B1, A1B1, A1B2, etc.), then you
need to select the between-subject error term. This will be error term
1. If you are comparing levels within a within-subject factor, you
need to select the error term associated with that factor. For
example, if you are comparing A1 to A2, then use the error term
associated with A. All between-subject comparisons will always use
error term 1. Keep in mind, group*within-subject factor will be a
within-subject effect and should use the error term of the
within-subject factor.
Does that make the selection of the error term easier?
>
> GLM-flex automatically chooses Error Term 4 in the proposed case, which seems to match the analogue one-sample T-test really well.
> In fact, all ETs 2,3,4 look similar, while ET=1 yields way more conservative results than said T-test in my case.
As stated on the webpage, the automatic approach might not always
work. This is a clear case where there are problems with how it
chooses the error terms. As a result of this and other cases, we are
working on removing the automatic error term selection and will soon
require that it be explicitly specified by the user.
ETs 2,3, and 4 are within-subject error terms.
ET 1 is the between-subject error term.
>
> With kind regards,
> Johannes Keyser
> Institute of Cognitive Science
> University of Osnabrück
|