I have a longitudinal dataset where I am looking to examine whether individuals change their category membership of a variable (with four categories) over four time points (baseline, 1y, 2y, 3y). Ideally I would like to see how individuals change this category membership across all of the time points, not just if they do. I've looked at a number of different tests that could be used for this, but short of creating some sort of change metric for each time point (which is not ideal as the categories are not ordinal), I am not sure how to do this.
Any suggestions welcome, thank you!
I am not entirely sure what you are looking for. Do you want to predict group membership for each individual or do you want to assess if and how the distribution to the categories of your ordinal variable changes over time? Please tell us more about the design, the research hypotheses and what you are exactly looking, i.e. for what purpose do you need "a number", so to speak.
Thanks for your response. I'm looking to see if individuals change their group membership over time, so does Person A who starts in category 1 change to different categories across time points, and if so in what way? I would also like to see is if there are patterns amongst the cohort in terms of the individual transitions between the categories. So, are there trends in terms of which categories people may transition to across time? I feel like that will require more than one test/assessment of the data, and the first stage may just be graphing it out to see individual trajectories over time.
As I mentioned in the question, these categories are not ordinal (it's a new way of looking at an existing measure), so it's not like I can look at going up/down ranks.
Thanks in advance - I hope that provides a bit more information to go on!
just for Rainer Duesing (and my own) sake, you mean that the response would be the change in category. For example, you would have 60 test subjects (people) s={si, ..., s60} and 4 categories c={ci,....,c4} (without specific order). After year specific time periods x={baseline, 1y, ..., 3y} you want to see if the subject changed categories?
Thus, something like, subject 1 (s1) changes from c1 to c2, from 1y to 2y? Here the question comes ..., what is your independent (predictor) variable? The categories are your dependent variables, right? you want to "predict" the change in categories as a function of something c=function(something).
One could thingkof a model c=f(x)+s where your years are the independent variables and s the random effect, which is your subject. Then are you thinking about a multinomial GLMM? Giving a side eye to Rainer Duesing if this make sense how no experience with subject setups.
Thanks for this Wim Kaijser - I think this would be the next step. Yes, we are ultimately looking at seeing whether an IV would predict the change in category membership over time, but my initial first step would be looking to describe changes over time and if there are trends. I guess a bit like a repeated measures ANOVA for categorical data, where the outcome is group membership and the factor is time.
Rachel C Sumner sorry, I really misread your post that you dont have ordinal data. But I would agree with Wim Kaijser that a multinomial GLMM, as described, can handle your data, but I am not sure if this really answers your question. I have not enough experience with multinomial GLMM to exactly know what the output/results will give you. My initial question, which is also relevant to ordinal outcomes, is, what you want as "output". I really do not know if the multinomial GLMM gives you the "trends" for each subject individually (thats how I interpreted your initial request). Of course, you can get estimates for the change over time, for the whole outcome.
Unfortunately, I do not have any good resources, since this analysis is quite seldom in my field, ordinal analyses would be nicer ;-)
Is this a Markov model? You can express the data as transition probabilities from one state to another, and you can also test whether there are predictors of these probabilities. And whether it's a true Markov model, in which only the state at time t affects determines the probabilities of transition at t+1, or whether the state at previous times also affects the probability – the mover/stayer problem.
Hello Rachel C Sumner. If you tell us what software you use, people may be able to direct you to helpful examples of the types of models that have been suggested by Wim Kaijser, Rainer Duesing, and Ronán Michael Conroy.
Thanks Bruce Weaver - I mainly use SPSS, but I can also use Stata.
Thanks all for your time and thoughts on this!
I'll start with multilevel multinomial logit models.
For SPSS, take a look at the GENLINMIXED command.
- https://www.ibm.com/docs/en/spss-statistics/28.0.0?topic=reference-genlinmixed
The documentation for Target Options indicates that you can estimate a (nominal) multinomial logit model:
- https://www.ibm.com/docs/en/spss-statistics/28.0.0?topic=genlinmixed-target-options-subcommand-command
For Stata--at least for official Stata commands--I think you would have to use xtmlogit:
- https://www.stata.com/manuals/xtxtmlogit.pdf
Re Markov switching models, you could take a look at Stata's mswitch command:
- https://www.stata.com/manuals/tsmswitch.pdf
AFAICT, SPSS still does not have any very straightforward way to estimate that kind of model. Here's what I found:
- https://www.ibm.com/support/pages/can-spss-fit-markov-chain-models-transitions-across-states-categorical-variable
HTH.
- Badges
- Science topic
- Similar topics
- Psychology
- Cognitive Science
- Mind