I am working with panel data (5 waves of longitudinal data, total n=4000, there is missing data) which I am analyzing in cross-lagged panel models. I would like to compare several models to see how the temporal relation is between three variables. I have three variables of interest: V1 = dichotomous (y/n) V2 = continuous V3 = dichotomous (y/n) V3 is an irreversible disease status (thus: respondents can go from no to yes, but not from yes to no – they cannot recover from the disease). V1 and V2 are ‘normal’.
As V3 is irreversible, the status of V3 across the later waves is very similar for the majority of people, as only the respondents with new onset disease change. For example: from wave 1 to wave 2 and wave 2 to wave 3 there is some change, but from waves 3 to 4 and 4 to 5 the not enough respondents that go from no to yes for the program to estimate the associations for V3 (I get very high coefficients, which implies that V3 on wave 4 is almost a perfect predictor of V3 on wave 5). I have tried using the incidence (new cases, which is approximately 7% of the respondents at each wave) instead of the prevalence, but that does not solve the problem. (see next post)
1) We thought that maybe we should treat V3 as a time to event variable (DTSA). Is that possible in Mplus for cross-lagged panel models? 2) If not: Do you maybe have an alternative solution for our problem? 3) If DTSA is possible, do you have an example? In the paper: Discrete-Time Survival Mixture Analysis by Muthén and Katherine Masyn: ‘’Conventional discrete-time survival analysis is a special case within this framework where a single-class latent class analysis of event history indicators is performed.’’ I thought I could maybe do that too, however, I have some trouble understanding how the censoring works in a panel model. This variable is both X and Y, and therefore I do not know how to incorporate the survival analysis in the model. - In some models, V3 is only a Y variable. For V3 as a Y, it should be related to the waves: if someone was a yes on V3 at wave 2, that respondent should not be present in the analysis for V3 as an outcome on wave 3. Does a single-class latent class analysis account for that? - In some models, V3 is only an X variable, or both X and Y (when V1 on w1 predicts V3 on w2 and V3 on w2 predicts V2 on w3). If V3 is a predictor, what variable do I use? I can use the original V3 variables, however, then I still have the same problem of the high association over the waves for V3.
I hope you can give me some advice on these topics.
That's a methods research topic that I can respond to only partially. V3 as time-to-event sounds reasonable. In principle it can be combined with a cross-lagged model but the trick is how to conceptualize it in the first place.
The User's Guide has setups for basic discrete-time survival analysis. You may want to have a look at our Survival page: