Stata command equivalent to R lcmm (for joint latent class analysis)

[Russell Nye]

Hi all,

I am wishing to identify latent trajectories of symptom score, a continuous variable ranging from 0-100, measured across 2 years at 6 different time points (baseline, 2, 6, 12, 18, and 24 months) in ~600 patients. I have reason to believe there is informative censoring due to patient death, and would therefore like to account for this in the model. Research with a similar objective, measurement types, and censoring issue have used joint latent class analysis to overcome the issue of informative censoring.

These analyses have all been performed in the lcmm package in R (lcmm manual for reference), and I am wondering if there is a Stata command equivalent?

Also, if anyone is aware of any research using joint latent class analysis to account for informative censoring in Stata would you kindly direct me towards that research?

Thanks,
Russell

[Erik Ruzek]

This may be possible with Stata's gsem, although the notion of informative censoring suggests you want to model that rather than account for the censoring. Check to see if gsem has what you want (help gsem family and link options). If you want to model it somehow, I'm not sure gsem can handle that. Weiwen Ng has a lot of experience with latent class analysis and may have some insight.

[Weiwen Ng]

This may be possible with Stata's gsem, although the notion of informative censoring suggests you want to model that rather than account for the censoring. Check to see if gsem has what you want (help gsem family and link options). If you want to model it somehow, I'm not sure gsem can handle that. Weiwen Ng has a lot of experience with latent class analysis and may have some insight.

Thanks for the mention! I have not personally done this, but it sounds like the OP would want to Google: joint modeling of longitudinal and survival data. (I will subsequently say joint modeling for short.) Indeed, this happens in many randomized trials for health care outcomes. Currently, the field treats each outcome on its own. With low mortality rates, this is probably fine. However, when you get to older adults, particularly if they're sick, or if you're dealing with cancer patients or some other disease with high mortality rates, the mortality rates are high enough that in theory, you want to think about how you would model the survival process alongside the outcome of interest (e.g. symptoms, function). Stata doesn't have native commands that can do this, I think. There appears to be a 3rd party Stata package that does this type of modeling, and Yulia Marchenko has slides on it.

I'm skimming the documentation for the R package lcmm. I see a reference to "joint latent class mixed models for longitudinal and survival data", which I am guessing is the finite mixture model version of the above model type. In general, in an FMM, you assume that there are k latent classes, and that E(Y) = f(XB) in each latent class, and that the betas differ among classes. So, as if there are heterogeneous treatment effects. Right now, I know that the native gsem command can't handle both discrete (i.e. latent class) and continuous (i.e. random intercepts/slopes) latent variables in the same model. I'm not currently aware of a Stata package that handles a finite mixture of joint models.

I'll point out that in principle, the situation as described by the OP only makes a case for regular joint modeling. I think that most reviewers would regard this as more than sufficient. As with times when people ask about zero-inflated models, I would ask if you can make a theoretical case for why there is heterogeneity in the response. If you can't make a reasonable case as to why you think there is more than one average that's of interest, then I think you're probably OK going with just a regular joint model. Naturally, if Stata did implement flexible mixture of joint models, you could conduct exploratory analysis. Nothing wrong with that. But I don't think Stata has this.