Original issue/concern/discussion: incf-nidash/nidmresults-fsl#125 (comment)
Filing on behalf of the shy @mih who might otherwise remain silent ;)
Given that a choice of a "canonical" HRF is nothing but a sample from a specific population (cats?) and region (early visual cortex?), I would say that it actually should be a "healthy practice" to introduce into the model a specification of area-specific (more plausible, based on a good prior) model of the response. , if prior assumptions exist (not just to include result of "fishing expedition/p-hacking"). Some models might even include/account for non-neural responses (e.g. to account/regress out mean blood influx effect correlating most IIRC 3 seconds after, if we assume that it is too distant from the neural effect of interest) which would require convolution with some custom kernel.
But I guess this issue is not about discussing best or worst practices, but rather about addressing use-cases existing in the wild, and since @mih ran into one upon a first shot -- there is a good number of them, since the tool allows for them. FWIW I would strongly advocate for extending the model to allow such flexibility in design specification to
- make it applicable to real world cases (not the ones crafted specifically to be represented by NIDM-Results)
- make it capable of describing more "sophisticated" designs
- not restrict users and software developers seeking to support NIDM-Results into what might be "suboptimal" practices.
Original issue/concern/discussion: incf-nidash/nidmresults-fsl#125 (comment)
Filing on behalf of the shy @mih who might otherwise remain silent ;)
Given that a choice of a "canonical" HRF is nothing but a sample from a specific population (cats?) and region (early visual cortex?), I would say that it actually should be a "healthy practice" to introduce into the model a specification of area-specific (more plausible, based on a good prior) model of the response. , if prior assumptions exist (not just to include result of "fishing expedition/p-hacking"). Some models might even include/account for non-neural responses (e.g. to account/regress out mean blood influx effect correlating most IIRC 3 seconds after, if we assume that it is too distant from the neural effect of interest) which would require convolution with some custom kernel.
But I guess this issue is not about discussing best or worst practices, but rather about addressing use-cases existing in the wild, and since @mih ran into one upon a first shot -- there is a good number of them, since the tool allows for them. FWIW I would strongly advocate for extending the model to allow such flexibility in design specification to