Unexpected nodes generated in model

[eort]

Hi all,

I am trying to set up a hierarchical regression on multiple DDM parameters. In my last attempt, I came across a surprising collection of nodes when printing out arviz.summary on the model inference data.
Here is the model I set up:

model = hssm.HSSM(data=data, z=0.5, model='ddm', 
    loglik_kind='approx_differentiable',  extra_namespace={"lvl": cfg['drugs']}, hierarchical=True,
    include=[
                    {
                        "name": "v",
                        "formula":"v ~ 1 + (1|subj_idx) + coh*C(drug, levels=lvl)",
                        "link": "identity"
                    },
                    {
                        "name": "a",
                        "formula":"a ~ C(drug, levels=lvl)",
                        "link": "identity"
                    }

    ])

with this being the associated graph:

After the sampling, I checked the summary with arviz and got a table that has following row indices:

Critically, there are 60 of v_1|subj_idx[XX] nodes, which matches the number of subjects I have. However, for a it generated 90563 a[XXXXX] nodes, which is the number of trials I have in the dataset, which is not what I want.
It seems that by using the hierarchical flag, and not specifying which variable in the dataset to use as random factor, hssm uses all row index? Not sure whether this is intentional, but it appears to be a little odd. So, if you wanted a random effect only on one parameter, but not on another, that wouldn't be possible?

Thanks,
Eduard

[frankmj]

Hi Eduard, this is not actually an error in the model estimation -- it is not estimating 90K different parameters for a -- but rather it simply stores the computed values for each regressor on each observation once the regressors from that trial are taken into account (ie. the deterministic outputs of the regression). These are stored in the trace object so arviz displays them by default. You can simply pass an argument to arviz to exclude those outputs using e.g. var_names = ['~a'] Michael Michael J Frank, PhD | Edgar L. Marston Professor Director, Carney Center for Computational Brain Science <https://www.brown.edu/carney/ccbs> Laboratory of Neural Computation and Cognition <https://www.lnccbrown.com/> Brown University website <http://ski.clps.brown.edu>

…

On Fri, Oct 13, 2023 at 8:01 AM eort ***@***.***> wrote: Hi all, I am trying to set up a hierarchical regression on multiple DDM parameters. In my last attempt, I came across a surprising collection of nodes when printing out arviz.summary on the model inference data. Here is the model I set up: model = hssm.HSSM(data=data, z=0.5, model='ddm', loglik_kind='approx_differentiable', extra_namespace={"lvl": cfg['drugs']}, hierarchical=True, include=[ { "name": "v", "formula":"v ~ 1 + (1|subj_idx) + coh*C(drug, levels=lvl)", "link": "identity" }, { "name": "a", "formula":"a ~ C(drug, levels=lvl)", "link": "identity" } ]) with this being the associated graph: ![image]model_graph.pdf <https://github.com/lnccbrown/HSSM/files/12893110/model_graph.pdf> After the sampling, I checked the summary with arviz and got a table that has following row indices: [image: image] <https://user-images.githubusercontent.com/8819465/274893446-86d90bb3-dae0-49e8-a0e6-27f8e8272283.png> Critically, there are 60 of v_1|subj_idx[XX] nodes, which matches the number of subjects I have. However, for a it generated 90563 nodes, which is the number of trials I have in the dataset, which is not what I want. It seems that by using the hierarchical flag, and not specifying which variable in the dataset to use as random factor, hssm uses all row index? Not sure whether this is intentional, but it appears to be a little odd. So, if you wanted a random effect only on one parameter, but not on another, that wouldn't be possible? Thanks, Eduard — Reply to this email directly, view it on GitHub <#298>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAG7TFE7QG4J6NFF3BLVBTLX7EUR5AVCNFSM6AAAAAA57BL3WGVHI2DSMVQWIX3LMV43ERDJONRXK43TNFXW4OZVG4ZTENBVGI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[eort]

Ah ok, I see. But are those 60 subject wise v-nodes separate parameters? I guess I am bit confused why these two parameters (a and v) are treated differently then.
Is there a way to only select the fitted nodes so that I can check convergence, etc.? Or do I need to select them individually per model, based on what I specified?
Thanks in any case!

[frankmj]

THere are only 60 nodes v|subject because these are subject-wise intercepts. Technically there should also be deterministic v nodes for each trial based on your regression as well, separately from the subject wise intercepts. You should be able to look at r-hat for all the parameters that you want to look at by using var_names (to either include selected ones or exclude others). I think the development team is working on a way to try to make this less cumbersome so that it doesn't display the deterministics by default.

…

On Fri, Oct 13, 2023 at 8:40 AM eort ***@***.***> wrote: Ah ok, I see. But are those 60 subject wise v-nodes separate parameters? I guess I am bit confused why these two parameters (a and v) are treated differently then. Is there a way to only select the fitted nodes so that I can check convergence, etc.? Or do I need to select them individually per model, based on what I specified? Thanks in any case! — Reply to this email directly, view it on GitHub <#298 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAG7TFEWQACJQJHGR3VYA33X7EZFDAVCNFSM6AAAAAA57BL3WGVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM3TENZTGUZTC> . You are receiving this because you commented.Message ID: ***@***.***>

[eort]

Alright, thanks again!

[digicosmos86]

Additionally, the hierarchical knob here is intended to convert the entire model hierarchical. When it's set to True, every unspecified parameter will essentially have a random intercept according to participant_id column, but it will do nothing if participant_id column is not present in the data. Therefore, hierarchical setting is not doing anything here. We have give this some thought and will throw an error in the next version if participant_id is not present in the next version.

Please see the documentation here for more details

[eort]

Yes, thanks. I saw that. I tried later to replace my subj_idx with participant_id, and indeed random intercepts are added to all three parameters (a, t, v). Interestingly though, those deterministics only appear for t and a, but never for v.

[digicosmos86]

Yeah this is a bambi-specific behavior. We have communicated this with the bambi developers and they are working on a fix. Please stay tuned :)