Thanks for the clear explanation. That makes a lot of sense.
Quoting "MCLAREN, Donald" <[log in to unmask]>:
> Hi Colin,
>
> Yes. The interaction contrast is still valid [0 0 0 0 1 -1 -1 1].
>
> The only difference between the F- and t-contrasts when there is only 1 row
> is that the F-test does not have a direction associated with it. The
> F-value is t^2. The significance of the F-test is the same as the two-sided
> significance of the t-test.
>
> You can determine if your contrast is testing what you expect by starting
> with the null hypothesis and using the null hypothesis to form the contrast.
>
> Multi-row F-contrasts are needed when you have multiple equalities in your
> contrast (e.g. H0: A=B=C).
>
> Best Regards,
> Donald McLaren, PhD
>
>
>
> On Mon, Oct 8, 2018 at 3:27 PM <[log in to unmask]> wrote:
>
>> Thanks Donald.
>>
>> I don't think the main effect of group is very sensible anyways, since
>> it is collapsing between time points and therefor of little interest.
>>
>> The interaction ioss till valid though right?
>>
>> We were planning a contrast os 0 0 0 0 1 -1 - 1 (the 4 interaction
>> terms, skipping the main effects terms) as an F contrast. I'm not 100%
>> sue that is right though. I'm not very experienced with F contrasts.
>>
>> I'd appreciate a second opinion : )
>>
>> Colin
>>
>>
>> Quoting "MCLAREN, Donald" <[log in to unmask]>:
>>
>> > Hi Colin,
>> >
>> > (1) You need to add the subject factor as a main effect. Then it will
>> > appear in your model.
>> >
>> > (2) The main effect of group is not a valid contrast in a repeated
>> measures
>> > design. This is due to the wrong degrees of freedom and the wrong error
>> > term. The error term of this model is the within-subjects error. For the
>> > main effect of group, you'd want the between-subjects error term, which
>> is
>> > not provided with the model.
>> >
>> > Best Regards,
>> > Donald McLaren, PhD
>> >
>> >
>> >
>> > On Fri, Oct 5, 2018 at 4:46 PM Colin Hawco <[log in to unmask]>
>> wrote:
>> >
>> >> Oh and I forgot part 2, my contrasts. Main effects are easy (1 1 -1 -1
>> >> or 1 -1 1 -)
>> >>
>> >> but for interaction, since it put the interaction terms in the model,
>> >> I think it would be an F contrast of [0 0 0 0 1 -1 1 -1]
>> >> Confirmation of this would make me feel a lot better, I've been as
>> >> confident in f contrasts as I'd like as I so rarely make use of them!
>> >>
>> >> best,
>> >> Colin
>> >>
>> >> Quoting [log in to unmask]:
>> >>
>> >> > Dear all,
>> >> >
>> >> > I'm sure this has been addressed before but my list search didn't
>> >> > run up a clear answer (a reflection on my poor search skills than
>> >> > the clarity of past answers, I am sure).
>> >> >
>> >> > I am running a repeated measures type ANOVA design, with a group
>> >> > (between subject) by time/session (pre-post, within subject) design.
>> >> > After some consideration, flexible factorial seemed the best way to
>> >> > go.
>> >> >
>> >> > I set for main effects of time and session, as well as the
>> >> > interaction. My design matrix is attached.
>> >> >
>> >> > I set independence for no for 'time', but not for group, while I
>> >> > left variance unequal (after all we expect changes over time, so I
>> >> > expect possible unequal variance).
>> >> >
>> >> > First I wanted to check if this seems OK.
>> >> >
>> >> > Second, I wanted to check if maybe we should model subject as an
>> >> > additional factor? It seems to maybe be already embedded implicitly
>> >> > in the flexible factorial. If I add this factor, but don't specify a
>> >> > main effect, it doesn't appear in the design matrix, which I found a
>> >> > bit surprising (it should still be modeled even if we don't contrast
>> >> > it, for the effects on the Beta estimation).
>> >> >
>> >> > Thanks a lot,
>> >> > Colin
>> >>
>>
>>
>>
>>
|