suggestions for fitting tvmvar / tvmgm to a group level dataset

[kamione]

I have read the mgm paper and tutorial paper for tvmvar. I am wondering if these methods are only available for single subject data? If not, how can I fit my data to the model. If yes, any suggestions for fitting my group-wise data? I thought about first fit the data individually and then combine all the graph to a group graph. Any thought? Thank you so much.

[bernard-liew]

Hi Jonas,

Yes I too am interested in applying time-varying graphical methods to group data, something like the mlVar package. Do you have any advise?

Regards,
Bernard

[jmbh]

Hi both,

Sorry for the late reply!

Certainly you could fit the individual models and then inspect/summarize the parameter estimates across individuals. However, this comes with all the issues of a stepwise approach. For example, when modeling the differences across individuals you don't consider the uncertainty about the estimates in the N=1 models. Also, you don't have the shrinkage towards the grand mean of parameters, which you have in standard multilevel models with Gaussian latent variable modeling the heterogeneity across individuals.

To build a joint model, I would start with a multilevel VAR model and then add parameters to allow for time-varying parameters. This could just be a main effect of time, but also interaction effects between lagged variables and time. Depending on how many subjects and time points you have, you might be able to include all the interaction effects. A first step would be to only include linear moderation, but of course you could in principle also extend this to a GAM, similar to the N=1 time-varying VAR implementation of Bringman et a.l (see, for example here: https://www.tandfonline.com/doi/full/10.1080/00273171.2020.1743630). Since VAR models can be estimated equation by equation, you could implement this yourself relatively easily with standard multilevel packages such as lme4 or brms.

On a conceptual note, when estimating time-varying effects in a grouped setting, the environments of the individuals of course have to be aligned for the analysis to make sense.

Best wishes,
Jonas