Hi, I am new to Mplus and trying to explore its capabilities to analyze some ecological momentary assessment type data. I have been conceptualizing as multilevel data - where individuals (level 2) have completed multiple (18) repeated measures of stress and positive affect over a 6 week time span (level 1), each of these measures are made up of 4 items. We have two health outcome variables that were measured at the end of the study.
I am interested in to what extent do stress and positive affect scores (the level 1 variables) predict these health outcomes (the level 2 variables measured at one time point)?
For this I could have created a mean score for level 1 variables and just done single level multiple regression in SPSS etc. - but I was hoping to make better use of the repeated nature of the level 1 variables. Is this something I can do in Mplus without separately aggregating - and if so how would this be specified - (while addressing any auto-correlation issues)?
Many thanks in advance for any assistance in this.
Let me first get clearer on the repeated measures. At how many time points are the subjects measured? And, how many measures are made at each time point? You say 18 repeated measures - is that 18 per time point (or 3 times a week for the 6 weeks)?
So our participants completed measures at 18 time points. 3 times a week for the 6 weeks. Approx. 150 participants.
For simplicity, in the original post, I just mentioned 2 things we measured at each time point - Stress and Positive Affect, each of which were short form questionnaires each containing 4 items. So 8 observed variables at each time point - although I can use the scales usual procedures to make just one score for stress and one score for positive affect at each time point if that simplifies things, as they are previously validated scales.
In reality we measured a few more things at each time point (e.g., negative affect, physical activity performance), so we have around 25 observed variables measuring 6 constructs at each time point. I was hoping that adding additional factors beyond the first two would be relatively simple - once I understood how to do this with just stress and positive affect.
The three days each week were chosen at random for each participant (but are consecutive). So no not all the same (although I could easily treat them that way as my data set is currently formatted in wide format so that I have variables for "week 1 first measurement-point" "week 1 second measurement-point" etc. But I also have variables (in SPSS) with the date each measure was actually completed. As well, If we decided to aggregate week-by-week to give only 6 time points then these could be treated as the same time point (although this is not my preferred option).
Well I'm interested in two aspects really: (1) mean over time (2) variability/stability of stress and positive affect (i.e., do those with more stable stress profiles have better or worse outcomes than those who fluctuate across the study).
I am not expecting a systematic increase or decrease in stress or positive affect across this time period, although I would expect them to fluctuate - and for some to fluctuate more than others.
I am asking these questions because of our current development work on Mplus Version 8 which incorporates methods for intensive longitudinal data like yours. Until that is available I see two options.
One option is to take a single-level wide approach for the 18 time points where the stress and PA are parallel processes (18*2+2 columns of data). If there is no trend perhaps only a random intercept for each of the 2 processes can be used (I am using growth modeling terms here). Those random intercepts (which varies across subjects) can then predict the two distal health outcomes.
Another option is to go the route of Wang-Hamaker-Bergeman (2012) in Psych Methods where you get into time-series analysis of the repeated measures and let characteristics of that analysis predict the health outcomes (this can be done in Mplus V8 using 2-level time-series analysis). They talk about temporal dependency and amplitude of fluctuations as predictors.
Thank you for these ideas, really appreciate it. Great to hear version 8 will incorporate methods for intensive longitudinal data – do you have an estimated release date at this point?
With the first suggestion, if some participants by chance have some upwards or downwards trend over time, would this make using the random intercept problematic? Would I just have to examine the plots to see if this might be an issue?
If that is not an issue - would this single-level method be specified in Mplus as below?
If you have a trend you can use a growth model with both intercept and slope growth factors as you have specified. You can play with the time centering point (see our Topic 3 video and handout) to define the intercept factor at different time points.
Regarding V8, we don't like to make predictions of release times - it won't be before the summer.
Thank you for that information. I have only been using Mplus for around 3 weeks so this is all really helpful.
I have now had the chance to look at the Wang-Hamaker-Bergeman (2012) method paper you mentioned. Thank you for recommending this as this is exactly what I was looking to do! I'm glad to hear this will be introduced into Mplus V8 (I assume this will be the one-step approach they describe?). However, until V8 is released - can I clarify if (or how) the multiple step approach they describe can be performed in Mplus V7? Can Mplus calculate ISD^2 and individuals autocorrelations, MSSD and peform detrending? Or is it a case of using another package to do these things and then using Mplus to model whether those features predict outcomes?