Jon Waldron posted on Monday, September 16, 2019 - 6:30 am
Dear Drs Muthen,
I have a question about the use of two latent variables derived from the same participants, which I was wondering if you could help with.
I have data from a cross-sectional survey into nightlife and drug use. I have run an LCA using binary indicators of past 12 month drug use to obtain six classes of polydrug use. I now want to run a separate LCA using binary indicators of whether different harm reduction strategies were endorsed, to see if different harm reduction profiles are related to positive and negative experiences following drug use. I expect that the relationship between harm reduction profiles and experiences will also be in some way influenced by participants past 12 month drug use.
One way that I thought of to investigate this would be by examining the relationship, e.g. via regressions, between harm reduction profiles and experiences separately for each of the six previously identified classes of polydrug use. I would also like to see how the two classes relate to each other, for example by examining the proportions belonging to different harm reduction classes within each polydrug use class. However, I have not come across this previously in the literature, and wonder whether using two latent class variables in this way is problematic, for example in violating any LCA assumptions about independence, or other?
Many thanks in advance for any assistance you are able to provide.
Jon Heron posted on Monday, September 16, 2019 - 10:08 am
this is quite do-able.
the first decision for you to make is whether you are happy for your drug-use indicators to potentially distort the measurement model for harm reduction and vice versa.
If the answer is no, which I suspect it will be, then the approach is to fit a pair of unconditional models C_drugs and C_harm and then bring these together using a manual implementation of the bias-adjusted three-step method.
Jon Waldron posted on Tuesday, September 17, 2019 - 6:07 am
Many thanks for your replies, Jon and Bengt.
Bengt - thank you suggesting UG ex 7.14, that looks exactly what I need in terms of looking at how the two latent variables relate to each other. Jon is, however, correct that I do not want drug use to distort the model for harm reduction and vice versa - is this achieved by specifying that thresholds for binary drug use indicators vary only for polydrug use class and harm reduction indicators only for the harm reduction class, as in the example for "cu" class?
Jon - in order to examine associations between harm reduction classes and experiences as an outcome, are you suggesting an extension of the method shown in webnote 15 (section 3) to include an additional class (polydrug use)? If so, would it be possible to point me in the direction of some example syntax that uses this method with two classes (I can't seem to find anything on the internet other than LTA, which is clearly not appropriate in this instance)?
Q1: No, not completely because the cross-construct associations affect each construct's solution.
Q2: Actually, I think the LTA approach outlined in WN15 is applicable to this case.
Jon Waldron posted on Wednesday, September 18, 2019 - 7:19 am
Thanks again, Bengt.
The only reason I thought the LTA example in section 4 webnote 15 would be inappropriate is that it mentions the relationship between two classes collected at different points in time, whereas my data are cross-sectional. However, if you think that it is appropriate to utilise this approach on data collected at one time point then I will certainly proceed as you advise!