I have a national sample of 700 individuals measured on 28 occasions. Measures were taken about once every two weeks, covering a whole year. Data were collected both by self compiled diaries.
Individuals were instructed to record, for every measurement occasion, what they ate and what they drank out of a long list of food products and drinks along with other descriptors.
Within a measurement day, composition of meals and any in-between meals are recorded.
I am trying to analyze whether what is consumed at one meal (or in between meal) occasion within a given day is related to what is consumed during another consumption occasion on the same day. Also, some food categories are consumed very regurarly (like bread, meat, etc.), while others display seasonality (or special occasion) effects.
Basically I am attempting to model meal composition pattern in a day, however my intuition is that I should not just collapse all 28 measurement occasions in a “average” day.
Apart from few individual level variables (e.g. age, body mass index), all other variables are categorical (mostly binary)
I was thinking about using a growth mixture model approach, but I find myself wondering whether there is any growth process at all in meal composition patterns. Do you have any suggestions or reference on the type of modeling that I could use?
Greenbaum, P.E., Del Boca, F.K., Darkes, J., Wang, C. & Goldman, M.S. (2005). Variation in the drinking trajectories of freshman college students. Journal of Consulting and Clinical Psychology, 73, 229-238
The paper is very interesting and I now have an idea on how to handle spikes in my data.
I also gave some more thoughts to my modeling problem.
Suppose an individual had only 1 meal per day. Also suppose that meal composition could be represented by a binary multivariate vector. If one considers m food/drink categories, the m-variate 0/1 vector would represent meal composition at time t as consumed by individual i. From time t to t+1, the study of meal composition patterns could then be represented as a transition in a m-dimensional 0/1 matrix.
Do you think that in principle one could:
- model meal compositions as a multivariate 0/1 vector
- model an individual's meal consumption process as a HMM
- add a latent categorical variable to capture unobserved heterogeneity in meal composition trajectories ?
On a side note, I realize my issue looks like a discrete choice situation in a dynamic setting...