Hi Statlist,
I'm working with longitudinal data featuring repeated measurements at baseline (BL), follow-up 1 (FU-1), and follow-up 2 (FU-2). My objective is to explore the impact of changes in marital status from BL to FU-1 on changes in the frequency of fruit intake at FU-2, relative to its baseline frequency.
To represent the transition in marital status, I created a new variable (MT1) based on marital status at BL and FU-1. For instance, someone married at BL and remained married at FU-1 is coded as 0=remained married, serving as my reference. There are a total of 6 categories for this exposure.
Here are my questions:
- Concerning the outcome variable, there is a variable named "fruit" that indicates the frequency of fruit intake per day (e.g., 0.43 times/day). Due to its fractional components, categorizing it as a count variable poses challenges, and treating it as a continuous variable (e.g., expressing "I eat fruit 0.43 times/day") is not sensible. As a solution, I introduced another variable called "fruit-count," ranging from 0 to 10, to represent the daily frequency of fruit intake and eliminate fractional components. Specifically, I recoded values greater than or equal to 0 and less than 0.5 as 0, and values between 0.5 and 1.5 as 1, extending this recoding up to 10. I am currently uncertain about whether to consider the outcome as a continuous variable ("fruit") or a count variable ("fruit-count"). From a statistical perspective, what would you recommend?
- When treating the outcome as a continuous variable using "fruit2" (at FU-2) in the xtmixed command in STATA 18 in a wide form, the iterative optimization process exhibits the message "not concave." However, replacing person (entity_id) with Province as Level 2 in the command resolves the issue. What could be the problem with using entity_id as the second level? (see below)
xtmixed fruit2 if gender_01_fu1==0
xtmixed fruit2 time_months if gender_01_fu1==0 || entity_id:
Performing EM optimization:
Performing gradient-based optimization:
Iteration 0: Log likelihood = -17434.275 (not concave)
Iteration 1: Log likelihood = -17434.275 (not concave)
Iteration 2: Log likelihood = -17434.275 (not concave)
.
.
Then I stopped it since it was quite a while and did not converge.
3. I plan to run a multilevel model (Level 1 (time), Level 2 (individuals)), specifically a Multivariable mixed regression model, stratified by gender and conditioned on baseline outcome variables, age, change in BMI, and SES. I believe utilizing the wide form of data may be more suitable than the long form. Can Mixed multilevel regression be applied using a wide-form dataset? If not, in the scenario of using long-form data, how can I assess the impact of exposure changes on the outcome at FU-2 while accounting for the baseline outcome?
I would greatly appreciate it if you could kindly furnish me with the appropriate commands (for any model you recommend).
Thanks for taking the time in advance!
