Strange results with log_logit link settings

[abelowska]

Hi,

I'm fitting the hierarchical full DDM model on Flanker data. I'm using safe priors with the log_logit link setting and (default) blackbox likelihood function. I've pasted model specification below:

hssm.HSSM(
    data=data_df,
    model='full_ddm',
    include=[
        {
            "name": "a",
            "formula": "a ~ 1 + (1|participant_id)",
        },
        {
            "name": "v",
            "formula": "v ~ 1 + (1|participant_id)",
        },
        {
            "name": "t",
            "formula": "t ~ 1 + (1|participant_id)",
        },
        
    ],
    z=0.5,
    sz=0,
    sv=0,
    st=0,
    prior_settings="safe",
    link_settings = "log_logit"
)

Unfortunately, I've got strange results:

• non-decision time does not converge
• a_Intercept value is very close to 0 (looks like issue similar to: Bounds and priors for generalized logit link #346)
• random effects for a are negative

I'm quite surprised because it is not very complicated model and I successfully fitted it with Stan.

When I bounded tau to (0.05, 0.4) , trying to force chains to not explore super small tau-s, the results are quite similar:

• a_Intercept value is very close to 0
• random effects for a are negative
• random effects for t are negative

These results make me think that the estimates of a and t are not in probability space but rather log/logit space. Can you confirm that? And if so, can you recommend how to properly recover estimates of t and individual's t (and theoretical other regression parameters)?

Appreciate any help!

[AlexanderFengler]

Hi @abelowska,

Two things,

First, in fact, we figured that there was a bug, effectively messing up prior settings when using the log_logit links. We are working to fix that, aiming for release by the end of the weekend.

The chain starting points are very suboptimal for some of the parameters when using log_logit setting and some unnecessary parameter limits are maintained.

Second, indeed, when applying the log_logit setting, the parameters are in logit space, so in principle, for better interpretation you may want to push them through the corresponding sigmoid.

I will add an issue to include more convenience functionality concerning this case, since it seems that we are, for now, focusing more on the log_logit case than we a priori anticipated.

Best,
Alex

[abelowska]

Thank for the reply. I'm looking forward to the release!

[eort]

Hi @AlexanderFengler ,

Just to confirm, that fix has not been released yet? I observe the same issue as @abelowska.

for better interpretation you may want to push them through the corresponding sigmoid.

What exactly do you mean with "corresponding"?

Thanks!