Tom Booth posted on Thursday, January 18, 2018 - 10:15 am
I am looking to model change in some cognitive test scores over time (4 points), to explore whether there are distinct profiles of cognitive decline over time (classes) and to investigate whether the decline and/or class membership relates to disease stage. These tests can either be combined into a single score, or multiple growth models fit (one for each test) and the classes defined by the growth parameters of all tests.
Sadly, the disease is fatal so there is also MNAR where some form of missing data model also needs to be included.
I am struggling a little to picture all these elements combined in a single model - especially given this is a rare disease so N is not large.
Ultimately, my question is quite simple, do you think such a combined model is feasible and implementable in MPlus?
There are various options to simplify the model, but the above would represent an ideal scenario. Each individual element is clearly do-able, but the combination feels like it may be prohibitively complex.
Muthén, B., Asparouhov, T., Hunter, A. & Leuchter, A. (2011). Growth modeling with non-ignorable dropout: Alternative analyses of the STAR*D antidepressant trial. Psychological Methods, 16, 17-33. Click here to view Mplus outputs used in this paper. download paper contact first author show abstract
but of course a small N can present problems unless the model doesn't have very many parameters. So with several processes (one for each test) there can only be a few parameters that are test specific.
You can do a Monte Carlo study to learn more about the sample size requirements.
Tom Booth posted on Sunday, January 21, 2018 - 10:49 am
I have been working with this paper to set up the model based on the total score (an initial step before considering each test as an individual process). Not modelling different classes of decline (defined by the growth parameters) these work fine.
I think the suggestion of a simulation to check on sample size is likely sensible, though I think with ~n=200 and 4 time points, it is safe to say the simpler the model the better.