|
|
LCA with varying indicators |
|
Message/Author |
|
|
Hello again! I'm looking at predictors of medication adherence at 1 year following a hospitalization. There are 4 binary manifest variables denoting adherence to 4 different meds. I am considering a latent class model with covariate effects on class membership. So I think an appropriate approach is something like: variables: ... categorical = medA medB medC medD; classes = c(2); analysis: type = mixture; model: %overall% c#1 on x1 x2 x3 ...; However, the caveat is that patients may have been prescribed any subset of the 4 meds, so not all meds are applicable for all patients. Setting the N/A med variables to missing doesn't seem right because the missingness is structural and is different than, say, missing because we didn't get follow-up on the patient. In other software such as SAS I have handled similar situations by using "long" format data, with varying numbers of records (1 to 4) per patient, a single binary "adherence" outcome and a 4-level categorical variable indicating the med, and have used a two-level model clustering on patient. I'm uncertain if this is the best approach in Mplus and if so what the syntax would be. Appreciate your thoughts. |
|
|
Challenging setting. I would think that the long data arrangement with varying number of lines per person is the same as wide with missing - at least this is the case in growth modeling. Apart from that, the key modeling issue though is that non-adherence in taking a med is not the same as missing. Hopefully, you have data on which meds each person was supposed to take. And, is the latent class variable an adherence/non-adherence distinction? It could also be a take more/less meds distinction, or both. |
|
Back to top |
|
|