Dear David,
You can also go Bayesian and compare model evidence maps for non-nested
models, see:
http://onlinelibrary.wiley.com/doi/10.1002/hbm.20327/abstract
http://www.sciencedirect.com/science/article/pii/S105381190900963X
I've got the same concern than Donald about your specific comparison though.
Best regards,
Guillaume.
On 07/11/16 11:52, David Hofmann wrote:
> Hi Donald,
>
> thanks for the answer. Apart from comparing different models, is there a
> general way to look at the model fit of a GLM, e.g. like an R² value or
> a something similar? I also thought about looking at the residual files
> which can be created by spm, but I'm not sure if there is an
> established/accepted way of inspecting the residual maps.
>
> greetings
>
> 2016-11-06 18:18 GMT+01:00 MCLAREN, Donald <[log in to unmask]
> <mailto:[log in to unmask]>>:
>
> In general, you could compute the Akaike Information Criterion or
> another model fit metric (e.g. Bayesian Information Criterion) to
> determine which model fits the data the best.
>
> However, if I recall the methods correctly, one approach in the
> referenced paper requires a separate model for each trial. In this
> case, I'm not sure how you can compare model fits across methods as
> the methods would have different numbers of models for generating
> the betas.
>
> Best Regards,
> Donald McLaren, PhD
>
>
> On Sun, Nov 6, 2016 at 10:17 AM, David Hofmann
> <[log in to unmask] <mailto:[log in to unmask]>> wrote:
>
> Hi all,
>
> I'm trying to calculate a beta series correlation for a
> relatively rapid event-related design with different estimation
> methods (see
> https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3251697/
> <https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3251697/>). Since
> I'm not sure if these methods will work well, I want to compare
> them. What would be a good way to inspect the model fit or the
> quality of the modelling in spm?
>
> greetings
>
> David
>
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|