Modeling longitudinal biomarker profiles PreviousNext
Mplus Discussion > Growth Modeling of Longitudinal Data >
Message/Author
 Mattia Valente posted on Monday, November 11, 2013 - 5:48 am
Hi everyone,

I'm a medical PhD candidate in The Netherlands with a research focus on heart failure and an interest in longitudinal modeling approaches. I have a decent amount of experience with basic regression techniques and some experience with linear mixed effects modeling for longitudinal data, but my current research questions and some googling have pointed me in the direction of LCGMs and GMMs.

A little background on the dataset and research problem: we have a fairly large dataset (1500+ patients) with biomarker data at multiple points. We are interested in identifying biomarker patterns associated with differential response (based on categorial outcomes, some with time-to-event data: death, rehospitalization, worsening/improvement in clinical status). We have a total of 30 different markers, categorized into 4 to 5 domains. We suspect there will be a number of trajectory 'clusters' in these groups of markers, associated with different clinical and pathophysiological profiles.

I would appreciate any pointers in terms of key publications, books, terms to look for, or courses to attend.
 Bengt O. Muthen posted on Monday, November 11, 2013 - 8:13 am
It's a good question how to best approach this richness of measurements. Hard to advice without knowing more and getting deeper into it is not the purpose of Mplus Discussion. But generally speaking it seems like LCGM and GMM can play a role and even LCA when you don't have a good feeling for the shape of the growth curve. Your situation is more complex in that you want to study groups of curves. A first step might be to separately analyze all curves in a domain and see if similar trajectory classes emerge as expected. Question is how one combines these curves within domain a then how one handles several domains. Does one see these outcomes within a domain at a given time as as multiple indicators of a factor for which growth is considered, or is there a factor behind the set of curves? See also

http://www.statmodel.com/download/Grimm%202009-RHD-SecondOrder.pdf

The basic LCGA and GMM topics are covered by our Topic 6 handout and video with references given there; see also our Papers section under Growth Mixture Modeling.

Going in a different and more technical direction, functional data analysis considers a curve as the basic variable to be studied. See e.g. James Ramsey's book.
 Bengt O. Muthen posted on Monday, November 11, 2013 - 9:19 am
Adding to my multiple indicator comment, you may instead view the outcomes for the different curves within a domain as being influenced by the same latent class trajectory variable (in either LCA, LCGSA, or GMM format), while outcomes from different domains have different latent class variables.

For time-to-event modeling, see Papers, Survival Analysis on our website.
 Mattia Valente posted on Friday, November 15, 2013 - 1:21 am
Thanks for the response! Plenty of reading and watching to do, and maybe start playing around with the software and some simpler models to get a feel for how it all clicks together. It's an ambitious project to be sure, but one that I was despairing of resolving with 'simple' mixed effects modelling.

I'm hoping to find some folks at my institution with more experience in this field as well, see if they can help me along.
 Bengt O. Muthen posted on Friday, November 15, 2013 - 4:56 pm
Good topic; good luck with it.
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: