Hi Guillaume,
Thank you for the encouragement, I'm looking at what to do with the paper next but for now at least people in the field can provide some feedback on how useful they think the paper is.
My experience has been that often researchers want to be able to specify the full design at the 2nd-level, rather than the approach of creating effects at the 1st-level and then specifying simple models at the 2nd-level. This is particularly true in the case of mixed-measures containing within-subject factors with > 2 levels. The Henson & Penny chapter contains some rather confusing guidance on using Kronecker product rules to generate the various tests in partitioned-error models for M-way ANOVAs. I personally think this approach is more opaque than specifying multiple error-specific models at the 2nd-level, particularly as the interaction terms grow. It therefore doesn't surprise me that this doesn't appear to have caught-on with most users. It seems the desire is to be able to specify the complete design at the 2nd-level and then test and plot the various effects from within this design. This aligns with the models discussed in the statistical textbooks and, I would argue, is easier to understand conceptually compared with the Kronecker product approach. Unfortunately, when implemented in the GLM, this approach does not align with the partitioned errors which I assume most people are after, and the leap into over-parameterised designs makes the contrast weights more unintuitive. This is where I believe a lot of the discussion, confusion and issues on the list have come from. To that end, the paper has been written for users who wish to represent the complete design at the 2nd-level (in a similar vein to tools such as MRM, SwE, 3dMVM and GLM_FLEX) whilst also retaining tests using partitioned errors.
Thanks for pointing towards the Wiki, I must admit I haven't looked at it for a while. Will's explanations are nice, but again this is based on contrasts of effects at the 1st level, rather than the full factorial design at the 2nd-level. The same is true of Rik's chapter, which continues to push the confusing Kronecker product rules. I think you may be on to something with a specific option in the batch and it would be great to see this included in a future release. There is clearly a desire for something like this from the community. Ultimately, my concern is not with the approach taken (whether it is via the Kronecker rules, error-specific 2nd-level models or pooled-errors) but simply that researchers are correctly specifying the models they think they are. I dare say there are many instances of this having been done incorrectly (for instance, the Gläscher & Gitelman tutorial includes subject effects, a la partitioned errors, but then gives weights that depend on the group size for testing all effects in the same model, a la pooled errors). A suitable option in SPM for this would make a huge difference I feel, although I realise these things take time!
Best wishes
Martyn
-------------------------------------------------
Martyn McFarquhar, PhD
Lecturer in Neuroimaging
G30 Zochonis Building
The University of Manchester
Brunswick Street
Manchester
M13 9GB
+44 (0)161 306 0450
-------------------------------------------------
On 11/04/2018, 17:16, "Guillaume Flandin" <[log in to unmask]> wrote:
Dear Martyn,
Thanks for sharing your paper, that was quick - I'm sure it could be
published as a technical note in another journal, at least you have a
DOI now (no, I was not a reviewer).
Our usual citation is the Henson&Penny book chapter you refer to:
http://www.fil.ion.ucl.ac.uk/~wpenny/publications/rik_anova.pdf
and Will wrote more recently some practical examples on how to implement
the partitioned error approach:
https://en.wikibooks.org/wiki/SPM/Group_Analysis
There's room for improvements though and feedbacks are welcome on what
is missing/unclear.
The code attached to the post below was also meant to facilitate the
practicalities of specifying several models by introducing a new option
in the batch interface for partitioned models:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=spm;7077f17d.1603
Rik's more recent book chapter is also an interesting read:
http://www.mrc-cbu.cam.ac.uk/wp-content/uploads/2015/03/Henson_EN_15_ANOVA.pdf
Now, for Karina's question: it's a 2x2 within-subject design so
everything can be tested with a simple one-sample t-test. Form the
contrasts of interest at the first level ([1 1 1 1], [1 1 -1 -1], [1 -1
1 -1] and [1 -1 -1 1]) and enter each of them in a separate one-sample
t-test. That's the partitioned error approach. For the pooled error
approach, you would proceed with the flexible factorial design and
specify a main effect of factor 1 and an interaction with factors 2 and
3. The contrast for the time by self interaction would then be [1 -1 -1
1 0 ... 0].
Best regards,
Guillaume.
On 11/04/18 12:14, Martyn Mcfarquhar wrote:
> Dear Samantha,
>
>
>
> Thank you very much. The paper is now available as a preprint on
> PsyArXiv for anyone who wishes to read it: https://psyarxiv.com/a5469
>
>
>
> Best wishes
>
> Martyn
>
>
>
> -------------------------------------------------
>
> Martyn McFarquhar, PhD
>
> Lecturer in Neuroimaging
>
> G30 Zochonis Building
>
> The University of Manchester
>
> Brunswick Street
>
> Manchester
>
> M13 9GB
>
>
>
> +44 (0)161 306 0450
>
> -------------------------------------------------
>
>
>
>
>
> *From: *Samantha Brooks <[log in to unmask]>
> *Date: *Wednesday, 11 April 2018 at 10:10
> *To: *Martyn Mcfarquhar <[log in to unmask]>
> *Cc: *"[log in to unmask]" <[log in to unmask]>
> *Subject: *Re: [SPM] 2 x 2 Repeated measures ANOVA
>
>
>
> Dear Martyn,
>
>
>
> Your aforementioned paper would indeed be an asset to the imaging
> community - we too are having various discussions about modeling a
> complex ANOVA at first level in SPM.
>
>
>
> Looking forward to reading your paper,
>
>
>
> Best regards,
>
>
> Samantha
>
>
> _______________
>
> Dr Samantha Brooks, Ph.D (Download
> Website)<http://www.drsamanthabrooks.com/>
>
> Dept. of Psychiatry,
>
> J2 Building, Groote Schuur Hospital
>
> Anzio Road
>
> Observatory
>
> Cape Town
>
>
>
> On Wed, Apr 11, 2018 at 10:51 AM, Martyn McFarquhar
> <[log in to unmask]<mailto:[log in to unmask]>>
> wrote:
>
> Hi Karina,
>
> Adding the main effects and interactions in the flexible factorial
> module only adds these effects to the design matrix, it doesn't
> create the contrasts for those effects. I'll discuss creating the
> contrasts separately below, but in terms of your design being a 2 x
> 2 fully within-subject, you will need multiple models to calculate
> the F-statistics using the correct error term. For instance, your
> effects are:
>
> Main effect of time: needs a model containing Time and Subject after
> averaging over Self
> Main effect of self: needs a model containing Self and Subject after
> averaging over Time
> Time x Self: needs a model containing Time, Self, Time x Self,
> Subject, Time x Subject and Self x Subject
>
> The contrasts are another matter because of the necessity of
> creating estimable functions of the parameters in overparameterised
> designs. In unbalanced designs this is particularly tricky given
> that most people are after the standard Type III sums-of-squares.
>
> Because this is such a common issue on the list, I have actually
> written a paper going over the how and the why of both of these
> issues, with an example in SPM. I submitted it to NeuroImage, but
> the reviewers came back stating that none of it was new and no one
> would find it useful. Not that I'm bitter, but it'd be nice to
> actually know that some people would find it useful! I'm planning to
> put it up as a preprint on arXiv, but I'll send you the current
> draft separately. I hope it will help answer your questions and help
> you understand what you need to do.
>
> Best wishes
> Martyn
>
> -------------------------------------------------
> Martyn McFarquhar, PhD
> Lecturer in Neuroimaging
> G30 Zochonis Building
> The University of Manchester
> Brunswick Street
> Manchester
> M13 9GB
>
> +44 (0)161 306 0450
> -------------------------------------------------
>
>
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|