Need help with figuring out setup and direction using HDDM/HSSM

[AbiGY3]

Hi all,

To preface I'm a psychology major with mostly self taught coding and modelling skills, so I apologize if this isn't the correct place to ask these questions but I've been struggling to wrap my head around continuing my modelling.

I'll outline below what I ideally want to do and what I've attempted so far and then what I've been stuck on.

So my plan was to use HDDM to do the following:

Drift Rate (single trial) = Latent variable

Jointly inferred from:

• That trial’s Reaction Time (RT) and Accuracy
• The subject’s EEG data from that trial
• Hierarchical Bayesian model constraints

Research Questions:

Q1: Does LPP amplitude predict evidence accumulation on a trial-by-trial basis?

Q2: Is this effect location-specific? (anterior vs. posterior)

Q3: Does depression symptom severity moderate this mapping?

Models:

Using hddm.HDDMRegressor() except for Null Model

M0: Null Model

M1: v ~ 1 + aLPP + pLPP (LPP predicts drift rate)

M2: v ~ 1 + CDI + aLPP + pLPP (CDI shifts baseline drift)

M3: v ~ 1 + CDI + aLPPCDI + pLPPCDI (CDI moderation)

M4: v ~ 1 + LPP_Diff*CDI (Location Contrast, LPP_Diff = pLPP - aLPP)

aLPP and pLPP are trial level Late Positive Potential ERP amplitudes and CDI is the just the total score out of 27.

---

Everything in terms of coding has gone well so far, including the install of HDDM, uploading data and running my null model.

I was trying to figure out how to do my regression models and I ran into information in one of the google groups that has me questioning if including a within-subjects measure and a between-subjects measure in hddm.HDDMRegressor() is even possible or statistically sound. I saw that this is possible in HSSM but as I was exploring doing it in HSSM I've gotten confused about how to choose the right priors, whether I should be doing parameter recovery, how and when to center a model and more.

To summarize my questions:

Is it possible to do what I set out to do in HDDM or do I need swap to HSSM?

What are the general steps I should follow to do this? i.e. 1. Null model and check convergence 2. Simulating data and parameter recovery etc..

How do I figure out what priors I should be using?

When and how should I center the model?

Any other advice or recommendations are welcome! I'm also happy to provide any additional information or clarify anything above as needed.
Thank you in advance!

Abi

[AlexanderFengler]

Hi @AbiGY3 ,

few comments on your question:

First, yes we definitely recommend trying this with HSSM at this point, however if you really want to stick to HDDM, check out the resources in this paper.

Second, with HSSM you can have within and between subject effects using the wilkinson syntax via the formulae package We have a host of tutorials in the documentation.

In essence with HSSM you are allowed to use syntax like: v ~ within + between + (within|participant_id). This was not directly possible with HDDM, where regressions were defined either globally or subject wise.

Third, concerning priors etc. some pragmatical advice for now: Start with noncentered=False and see whether the model samples. You can print the model (in multiple ways, please refer to the docs, but the simplest: print(my_model)) to inspect all prior choices that where taken. All priors you see, you can manually adjust upon model specification. While we try to provide sensible defaults, in many cases you will want to make some adjustments to e.g. not allow absurd-to-your-application parameter setting to occur. In general its a good idea to z-score your predictors wherever possible, which the default priors for continuous covariates assume to have happened. You can define a model with z-scored covariates of course, but if you do, I'll flag that it is most likely that you will not have optimal results with the basic default priors.

If you play around with this more, feel free to continue the discussion here with follow-up questions. We will try to help you iterate.

Best,
Alex