hddm.HDDM versus hddm.HDDMRegressor

[Dan Bang]

For hddm.HDDMRegressor(), are the regression coefficients treated as fixed effects only or also as random effects? In addition, are subject-specific intercepts added? Or, maybe put more precisely, are separate intercepts and coefficients fitted for each subject while/before modelling the group-level effect as in mixed-effects models? 

I’m asking because I’m using confidence reports as an independent variable; subjects will report their confidence in different ways (e.g., tendency to low versus high reports) and the predicted effect might be better estimated for some subjects than others. I’m trying to gain an intuition as for whether the regression analysis follows the same hierarchical principles as the DDM model [or if I should deal with those subject biases before performing the hddm regression analysis]. 

A maybe related question which might shed some light. If I have a condition with three levels, what is the difference between running hddm.HDDM where let’s say drift rate depends on the condition and running cddm.HDDMRegressor where drift rate is predicted by condition [assuming that the drift rate in fact depends on the condition in a linear manner]?

These issues weren’t entirely clear to me after reading paper/online documentation. I hope I didn’t miss anything obvious!

*I’ve posted a set of questions separately so as to make it easier for others to find answers in case they should have similar questions.

[Sam Mathias]

For hddm.HDDMRegressor(), are the regression coefficients treated as fixed effects only or also as random effects? In addition, are subject-specific intercepts added? Or, maybe put more precisely, are separate intercepts and coefficients fitted for each subject while/before modelling the group-level effect as in mixed-effects models?

Its best to avoid the terms "fixed" and "random" effects when talking about Bayesian models, as those terms tend to mean slightly different things in different contexts. See this: http://andrewgelman.com/2005/01/25/why_i_dont_use/

My understanding (Thomas, please correct me if I'm wrong) is that when you add a covariate to your regression model, e.g., "v ~ x", you create subject-specific coefficients, which are then estimated hierarchically (i.e., group-level mean and std nodes are also created, and the subject-level coefficients are assumed to be random variables drawn from a parent distribution defined by those nodes). Therefore there is no "intercept" per se. However, if you are creating a model with one or more within-subjects effects then you do create an intercept (or baseline) covariate per subject, but this seems to be different to what you are trying to do.

A maybe related question which might shed some light. If I have a condition with three levels, what is the difference between running hddm.HDDM where let’s say drift rate depends on the condition and running cddm.HDDMRegressor where drift rate is predicted by condition [assuming that the drift rate in fact depends on the condition in a linear manner]?

Those are two completely different models. For the former, you are estimating the DDM parameters completely separately for your three conditions, as if there is no connection between them. This would make sense if you have three groups of subjects, as in this paper: http://brain.oxfordjournals.org/content/brain/early/2015/11/17/brain.awv331.full.pdf

In the second, you are imposing an order and magnitudes on your "conditions". Let's say "condition" is actually a value: 1, 2, or 4. By doing a regression, you assume that the ordering is meaningful and that the sizes of the differences between them are meaningful. If these are just different conditions in your experiment (A, B, or C), this is an inappropriate model design.

[Dan Bang]

Thanks for this!

--
You received this message because you are subscribed to a topic in the Google Groups "hddm-users" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/hddm-users/oSOJvQdb1f8/unsubscribe.
To unsubscribe from this group and all its topics, send an email to hddm-users+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

Dan Bang PhD

Postdoctoral Resarch Associate

Wellcome Trust Centre for Neuroimaging, UCL

https://sites.google.com/site/danbangphd/