LCA with covariates: is it still worthwhile to use 1 step approach?

Question

Dear statisticians' community, I am trying to compute a Latent Class Analysis through Stata and/or R. I built a 5 classes LCA model using poLCA on R and added a set of covariates.

It seems from the literature, like Vermunt (2010), that the use of 'One step approach' has some limits, but it's not possible to implement the 'Three steps approach' with the softwares that are available to me.

There's a mixed-R-Stata solution on the web, provided by Tompsett and De Stavola, but unfortunately the last step is a bit cumbersome and I do not understand how to transpose the example they report to another dataset (lost on the transpose-matrix-part, where var. 'n' is not specified by the authors).

So my question comes like this: could I just use 'One step approach', like suggested by B. Pratt (and cited literature) and obtain a valid result? Can I interpret the covariate coefficients or should I just use the covariates as a 'control', and just interpret the group item patterns? I am sticking to the covariate LCA model also because it has better AIC and BIC than the one without covariates. But I am scared that than I am obtaining statistical weak results.

I provide some code I used. The items are all dichotomous variables. N= 3,097

model without covariates:

f0 <- as.formula(cbind(life_proj, single, health, work_family, lifestyle, financial_hardship, no_support, climate_fut, economic_fut, child_impact) ~1)

LCA5 <- poLCA(f0, data=df, nclass=5, maxiter = 1500, nrep=6, graph=TRUE) #Log-likelihood; -18922.3; BIC(5): 38278.66

model with covariates:

f2 <- as.formula(cbind(life_proj, single, health, work_family, lifestyle, financial_hardship, no_support, climate_fut, economic_fut, child_impact)~hitdum+agedum+sex+rel+work+vivofigli+education+countrynu+area)

LCA5.cov <- poLCA(f2, data=df, nclass=5, maxiter = 1500, nrep=5) # maximum log-likelihood: -18394.3 ; AIC(5): 37040.59; BIC(5): 37801.4  questo è il migliore 

Thank you a lot for the help!

Irene

Answer 1

From my perspective, there is nothing wrong with using a 1-step approach in which the covariates are added directly to the LCA model to predict class membership. I would rather say the 3-step approach is often questionable because it implicitly assumes that the class solution will stay the same when covariates are added to the model--which is not necessarily the case.

I would check whether the class solutions (profiles of conditional item response probabilities) look the same with versus without covariates. If the profiles differ, this often means that the covariates have direct effects on specific items. This is worthwhile studying. The 3-step approach would ignore such direct covariate effects.