Use of estimated drift rate for fMRI modeling

[LeoCeravolo]

Dear all,
Thanks for this incredible tool !

I have a design as follows: [Population(15 Controls, 15 Patients)] * [Task(1 Explicit emotion categorization, 1 Implicit emo. categ.)] * [Voice Emotion(Happy, Angry, Neutral)]. The task was designed so that it was a 2-AFC (Implicit: female-male; Explicit: angry-neutral; happy-neutral). 5201 trials in total, so about ~173 per participant.
I am mostly interested in Population * Task, less about emotion here.
So I modelled my data (and sampled simulated data as well, with good fit) for each task including all participants and obtained--by a large margin, cf. r_hat and model comparisons--the best modeling with the Weibull distribution using LAN and the approx_diff likelyhood (numpyro). As expected/hypothesized, drift rate was higher for the Implicit than the Explicit task (~1.9 vs ~0.77).

Here is my question:
I have regions of interest but I also wanted to explore the trial-level correlates of the drift parameter to observe Population*Task brain modulations, is that possible (i.e., extracting trial-level drift rate with the method I used) or even recommended, or should I focus on using the lme4-style regression analysis in my ROIs (as implemented in HSSM) ?

Thanks a lot for your help, and sorry in advance if I forgot to mention some relevant aspects, I will happily add them.
{First time user, long time reader..}
Best,
Leo

[AlexanderFengler]

Hi @LeoCeravolo ,

To get the trial-level parameter values, try the add_likelihood_parameters_to_idata() function.
You can use it like: my_model.add_likelihood_parameters_to_idata().

If you check your trace after, you should now find included the trial level parameters for any parameters for which you had defined a regression model.

Let us know if this doesn't yield.

Best,
Alex

[LeoCeravolo]

Hi Alex,

Thanks a lot, it works and I have my values!!
{I noticed I did put a wrong number of draw in the mcmc, modified it according to my data, reran the inference and now it's fine}

Thanks again for your quick reply and for allowing me to move onto the final step of my 'plan'.

Cheers from Switzerland,
Leo

[LeoCeravolo]

Dear all,

I ran into a related issue but maybe I'm doing something wrong: when recovering the trial-level values, I cannot get the values for regressed parameters either for interactions or random slopes (less important), even though I can see them in the posterior traces of the model (see red squares in the attached image).
For instance if I want to get the values for the drift parameter, it works well.
However if I want to recover the drift as a function of my Task*Emotion interaction (v_Task:Emotion) of according to the random slope of Task per Population (v_Task|Population), I get an "unresolved reference" error.
It seems it's just a syntax thing, but have you ever encountered this and is there a solution?
Thanks a lot :)
Best,
Leo

[AlexanderFengler]

@LeoCeravolo,

sorry for the delay.
From what I see in the screenshot, the naming of these parameters is just not suitable to accessed in the way your are attempting it. Try instead to access the parameters via .posterior['name_of_parameter'].

[LeoCeravolo]

Hi Alexander,

It works thanks a lot, it was such a little syntax thing that I did not even think about it... my bad.
Thanks again,
Cheers,
Leo