I have ~35 days of zero-inflated substance use data and want to disaggregate between and within-person processes with a two-part model. I'm using RDSEM and including time trends for both the binary and continuous parts. I'm wondering...
1. Is it reasonable to have time trends for the zero-part and continuous part, and separate autoregressive parameters for both? I'm concerned I'm missing something, given Dr. Muthen's choice to model the autoregression with a single latent variable at the Johns Hopkins Short Course and his statement that two-part DSEM isn't quite there yet.
2. I read... somewhere... that RDSEM is not available for categorical data, but I'm able to estimate autoregressions for the residual of the zero-part of my outcome as long as I keep it as a fixed effect. Am I missing something?
3. None of my two-part models provide a DIC or any other fit criteria. Is there any empirical way to determine whether effects should be fixed or random?
Thanks so much for your response! Two quick follow-up questions:
1. My model runs in DSEM, but the same model in RDSEM gets the error "this model is not available. Variables regressed on lagged variables should also be lagged." The error is driven by the binary part. My within-person statement looks like this:
y^ on y^1; u^ on u^1; m1^ on m1^1; m2^ on m2^1;
y on m1 m2; u on m1 m2; ! commenting out this line makes the model run
Why is this a problem for the binary-part but not the continuous part? It seems consistent with other RDSEM code from papers/the website...
2. When I remove the offending lines, estimation blows up at the first iteration because the between-level posterior covariance matrix is NPD. I don't have this problem in DSEM, but I know from the appendix of Asparouhov et al. (2018) that DSEM will get point estimates correct but not variances, and so I wonder whether DSEM is overestimating the variances. Is this an indicator that I should simplify the model, or is there a better solution? (I'm running version 8.2).
1. The only thing that I can think of is that some of the variables have lag bigger than 1. All variables must have the same lag for this model. I would recommend using 8.3 and if that doesn't fix the problem send it to firstname.lastname@example.org. I recall some fixes regarding similar situations were done in 8.3
2. This problem is most likely data specific. To diagnose it make sure the two-level model without any autocorrelations is running ok. Also very small data sets could cause such a problem due to some autocrrelations >1. Also try this model first y^ on y^1; u^ on u^1; y on m1 m2; u on m1 m2; Again if unable to move forward send it to email@example.com