Hi
I'm looking at a longitudinal data set where people are followed up at different time points. In theory participants should be the same age at each time point but there can be up to 12 months variability. Also there is not always an equal gap in terms of time points.
I have two questions here, to measure change/growth in outcomes (using a mixed model) should I use time, time squared, time cubed etc. or age, age squared....?
Secondly do I need to control for the actual age, given that there is some variability from the age participants are supposed to be at the follow up (e.g. someone could be 5 yrs 6 months and someone else 5yrs 2 months, but they are both being followed up at age 5).
Thanks in advance
Tom.
I recommend actual age since you’re using a multilevel model.
i would avoid polynomials. Usually a log works well if the change is expected to reach an asymptote.
The actual age and time would be the best if that model passes diagnostic checks (normal residuals, etc.). If the model has issues you might need to consider transforms or using a correlation structure depending of the features of your data.
I too would use actual age (or even a polynomial of age if the response can decline with age and not just stop growing) centred around a convenient value of age to give an interpretable intercept. It really depends on the nature of what is being modelled but often there are growth spurts and you need fine detail age to get at that. I have also seen an example where truncated age does not find varying slopes while decimal age does - there is more information there; and more information means more power to detect effects. The multilevel (mixed) model's ability to model continuous time is an advantage over the SEM approach to growth models.
- Badges
- Science method