I have a analysis question that I was hoping to get some advice on. I am new to multilevel modeling so please point me elsewhere if you think it more appropriate. I want to make sure that this approach is appropriate before proceeding.
Each participant rates a set of 30 pictures for attractivienss (0-9 scale)2 times in the same assessment (Long-term mate vs. Short-term mate). Each picture differs by level of body fat (1-6) and muscle malss (1-5). The relationships between Attractivenss FAT & MUSC are quadratic. Thus a 3-D graph of attractiveness by fat & muscle would look like a mountain peeking at level 3 Fat & level 3 Muscle.
I'm interested whether GENDER, RACE/ETHNICITY, & RELATIONSHIP STATUS differentially predict the pattern of Attractiveness ratings for short term vs long term mates.
I am also interested in the amount of variability in attractiveness ratings explained by FAT & MUSC.
Could this be examined in a Multilevel model where GENDER, RACE/ETHNICITY, & RELATIONSHIP STATUS were level 2 predictors and ATTRATCTIVENESS as a function of FAT & MUSC were Level 1?
Thanks in advance for any thoughts?
bmuthen posted on Monday, February 20, 2006 - 6:35 pm
Instead of a multilevel approach, a multivariate approach is more flexible. This would use 30 variables per participant where you can freely model the relationships among the 30 variables. You need a large number of participants, however, but that is true also for a 2-level approach.
Thank you for the advice. Would 300 be a large enough sample?
bmuthen posted on Tuesday, February 21, 2006 - 3:25 pm
It might be sufficient, but is on the small side.
Mari Kim posted on Thursday, January 14, 2010 - 9:17 am
I have some questions to clarify my models.
The study investigated, a) whether individuals’ mood change in real-time was predicted by eye fixation patterns to a series of highly arousing negative stimuli, and b) whether this relationship varied by age and attentional functioning. They reported their mood ten times after viewing 12 images at each time point. Fixation is the percent of time participants spent looking at the most negative areas of the images during each time interval.
1) How many levels should be needed? There are 10 time points for mood recording and eye fixations data are nested within each time point. But we're interested in how the mean percent fixations at each time point predict individuals' mood. Can we include both Time and Mean Fixation as Level 1 variables, or should Fixation be Level 1 and Time needs to be Level?
2) If a two-level model works for this data, is it okay to include all the 117 fixation data for each individual. Or only 9 mean fixation (no fixation data for the first occasion) data for each individual needs to be included?
3)How Time variable should be modeled? Time points indicate when people report their current mood as they're viewing unpleasant images. Should we look at effect of time with respect to both the means and the variances? Or Is Time variable just to be included to control for fatigue effects?
You could average over the twelve values of mood and fixation and do a parallel process growth model of the two processes. This would be a single-level analysis if your data are in the wide multivariate format. You could also analyze it as a two-level model if your data are in a long univariate format.
Mari Kim posted on Friday, January 15, 2010 - 12:54 pm
Thank you for your answer Linda. I'm not familiar with the parallel process growth model you mentioned. In my model, I entered Mood as DV to predict its change from fixation patterns. How is it different from the MLM model? Also, I'm not clear if I need to enter a Time variable (1-10points of mood rating occasion) to predict Mood. Should I need a Time variable in my case?
Any thought is appreciated. Thanks a lot.
Mari Kim posted on Friday, January 15, 2010 - 12:59 pm
For a clarification, I only have 10 mood data for each individual. People rated their current mood before watching any image and then rated their mood occasionally after watching about 12 images. Hope this helps.
As a first step, you may want to simply regress Mood on fixation patterns (averaging over the 12 images to get one y and one x) using a two-level model where time is level 1 and subject is level 2. This can take correlation across time within individual into account by say a random intercept model. So that's not getting into growth modeling.
Mari Kim posted on Monday, January 18, 2010 - 7:15 am
That makes sense. So I enter Time (0-9 occasions) and Fixation as Level 1 predictors. When I only include 9 average fixations, the fixation didn't predict change in mood, but when I included all the fixation data with the average fixations, then it predicted mood significantly. I think it's because there were more data points by including all the fixation. Here, I wasn't sure if I only need to include the average fixations in the data file.