To use stimulus coding or not ?

[lauraglobig]

Hi All,

Thanks so much for the detailed documentation.
I am trying to determine whether I should be using stimulus coding and how. In my task participants make an initial choice, then get advice from 1 of 3 different advisors of differing accuracy, and then provide an updated choice.

I want to model the decision to conform or not to conform, and determine if the advisor type and accuracy influence the starting point or drift rate. I have recoded their initial choice, as whether or not it is the same as the advisor advice.

My data at the moment takes the following form

• Response : 1=Conform or -1 =Non-Conform
• Advisor: A,B,C
• Average Advisor Accuracy: -1=random, 1= accurate
• initial_agree: 1=Yes, -1=No

On top of other effects I am primarily interested in the effect of advisor type on starting point and drift rate, such that participants may be biased towards conform if the advisor is of one particular type, and that they also have a larger drift rate towards conforming vs non-conforming for one particular advisor type.

Am I right in thinking I should stimulus code initial_agree or should I also transform Accuracy, since participants are more likely to conform with accurate advisors? Would the following model set up and numeric assignment of the variables (see above) be correct?

df_hssm["initialdisagree.stimcoding"] = (df_hssm["initial_agree"] == -1).astype(int)


m_stim_inc_z = hssm.HSSM(
    data=df_hssm,
    model="ddm",
    include=[
        {
            "name": "v",
            "formula": "v ~ 1 + initial_agree+C(advisor)+accuracy",
        },
        {
            "name": "z",
            "formula": "z ~ 1 + initialagree.stimcoding + offset(initialdisagree.stimcoding)+C(advisor)"
        },
    ],
)

m_stim_inc_z.sample(sampler="mcmc", tune=500, draws=500) # set low for testing purposes

m_stim_inc_z.summary()

plan on testing interaction effects e.g. of accuracy and advisor later on, but wanted to focus on the problem at hand first.

Also - is there a way in HSSM to get the DIC for each model, as previously possible in HDDM?

Any help would be greatly appreciated!

Thank you!

[gpagnier]

Stim coding is appropriate. Your model looks ok but looking briefly at the stim coding tutorial, your covariates for predicting z may have to be refactored to initialagree.stimcoding = (2 * (df_hssm["initial_agree"] == 1) - 1) while initialdisagree.stimcoding should be (df_hssm["initial_agree"] == 0). Your drift rate regression equation looks good to me.

Also this model isn't hierarchical, you may want to allow parameters to vary by individuals depending on your question (the exact syntax will depend if you use a centered or non centered model).

It's possible to use pymc syntax to calculate information criteria (the core of the HSSM model is a pymc model after all) but a better way to evaluate your model is to use posterior predictive checks (which HSSM has lots of functionality for).

[lauraglobig]

Thanks so much! Yes I plan on adopting a hierarchical model. I just wanted to keep it simplistic early on.