Combining continuous and dichotom variables

[max-kat]

Dear community,

I am stuck on a problem for quite some time now and would really appreciate your thoughts on this:

My experiment has a 2x2 design. Participants have to imagine stimuli (condition A) and are shown stimuli (condition B). These stimuli consist of 2 types: Faces and Artworks.
After each presentation participants judged the evoked experience on how 1. pleasurable, 2. beautiful, and 3 moving it was.
Furthermore, only in the imagery condition they also evaluate how vivid the imagined picture was.

My research question is whether the neuronal representations underlying the elicited aesthetic experiences (pleasure/beauty/moving) across the different conditions (imagery vs perception) and stimulus types (faces vs artworks) are similar (or not).

In the first step, I created the single trial beta maps for each condition. In the next step I created a candidate model for perception vs. imagery and one for face vs artwork. These are dichotomous and easy to generate.
But then it gets difficult. If I just focus on one experience (e.g. moving) the candidate model is no longer dichotomous but now continuous (between 1-7). However, I guess I just use the raw values and then create a dissimilarity matrix. However, I am not sure how I combine all 3 ratings into one RDM that I could test. Could someone explain to me how this is done?
Furthermore, I have the suspicion that elicited experiences become more similar between perception and imagery when vividness was high and get more dissimilar when they are not very vivid. However, since vividness was only assessed for imagery, I have no clue how to incorporate this assumption into an RDM cause I only have values for imagery and none for perception.
I would really appreciate any help on this!
Also if you have time and really want to work with me on this, I'd be open for a collaboration!

Thanks for reading this far!

All the best,
Max

[JasperVanDenBosch]

Hi Max,

Furthermore, I have the suspicion that elicited experiences become more similar between perception and imagery when vividness was high and get more dissimilar when they are not very vivid. However, since vividness was only assessed for imagery, I have no clue how to incorporate this assumption into an RDM cause I only have values for imagery and none for perception.
I would really appreciate any help on this!

For this maybe you could split the data for imagery into several RDMs according to vividness, e.g. a low-, mid and high- vividness imagery RDM. (Take the low-vididness trials and create an RDM from that, etc) You could then compare the perception RDM with these one by one?

However, I am not sure how I combine all 3 ratings into one RDM that I could test. Could someone explain to me how this is done?

Here I think it depends on exactly what you want to hypothesise. Have you considered a MixedModel? You could either set static weights, e.g, task * 0.33 + moving * 0.33 + face vs art 0.33, or even have rsatoolbox determine flexible weights when fitting the model.

Hope this helps!

[max-kat]

[email protected] thanks for connecting me!!!

[max-kat]

one last question regarding your previous answer:

If you can't make a complete (abcd) RDM from say, a sub-group of 1/5 of your participants, I don't think this approach is the best one to take.

Do you mean that if I would need more than 1/5 of participants to form a full RDM, this approach wouldn't be great?
So, if I just use groups of 2 participants (2/33), it would be ok? I would just match them based on what they miss & have, so I would need only 2 participants on average to form a complete rdm. since some have complete cases as well I would roughly double my 'available' data. would that be a somewhat sound approach?

[JasperVanDenBosch]

Exactly. With 2/33 you'd end up with 16 RDMs which is reasonable to get some measure of variance. If you only had say, 4 RDMs that would kind of defeat the purpose. Note that you'd not be able to generalise across the population, since these RDM's do not stand for individuals but rather mini-groups. But at least you'd have some sense of the variance. You could also resample the participant combinations.

[max-kat]

Sorry, @JasperVanDenBosch one last conceptual question:

If I have 2 runs and a 2x2 design and all conditions are present across both runs, I can use the crossnobis distance to estimate similarities.
However, what if participants also rated the stimuli afterward (e.g. high & low for simplicity)?
Now let's say all high trials were in run 1 and all low trials were in run 2. Is it still 'valid' to use crossnobis distance? Cause for me - at least theoretically- the design changed from 2x2 across 2 runs to a 2x2x2 design (the last 2 are high and low) and thus crossnobis can no longer be used since high and low are not present across both runs. More specifically, I suspect that crossnobis might consider the evoked experience across runs now as noise (as the experiences are not present across runs to validate) and thus similarity estimates for the evoked experience would be attenuated. Is this a valid fear? Or can I still use crossnobis since the design initially was a 2x2 design?

[JasperVanDenBosch]

In this case you wouldn't have runs on which to crossvalidate so crossnobis wouldn't make sense (it would be mahalanobis).

PS Can you create new topics for subsequent questions? That makes it easier for other users to find their answers :)