I'm conducting LCGA to explore whether some classes in my sample are different on distal outcome.
In preliminary analysis I've identified one significant covariate likely to impact my distal outcome and I'd like to include it in comparisons tests between classes.
On the basis on the recommandation of Jung & Wickrama (2008), the first step is to choose the best-fitting model (based on fit indexes), by performing a series of models with an increasing number of classes.
I'm wondering if it's right to perform these series of models including the covariate. Or should I perform a series of models with an increasing number of classes without covariate, choose the best-fitting model with X classes, and after perform the chosen model with X classes and include at this moment the covariate in the model to explore difference between classes on distal outcome?
Does anybody have any suggestion? Or references using this type of data analysis with covariate?
Many thanks
Your reference seems to be out of date. I suggest that you repeat the attached screenshot search and carefully consider the more modern methods. Best wishes David Booth
I'm assuming you are doing growth mixture modeling or latent class growth analysis. In either case, it seems more common to solve the class enumeration problem first (without covariates), then introduce covariates to predict class membership in the class solution that you identified as best fitting. However, there is no fixed rule about this to my knowledge, meaning you could also include covariates right away. However, it seems less complicated to first identify the proper number of classes before introducing covariates of class membership.
Agree with Christian Geiser, there are no fixed rules.
Deciding the best number of classes (clusters) will become more complicated in the 1-step model (including covariates), as the global fitting index is determined by two factors (number of clusters and covariates included). If the 4-class 5 covariates solution is comparable with the 5-class 4 covariates solution, then which one is better? How do you choose covariates? The 1-step model also assumes conditional independence between observed variables and covariates, which can be easily violated. Another question is: how about the unobserved covariates. If only one out of three clusters can be predicted by observed covariates, will the model give more preferences to the two-class solution over the three-class solution?
There are some issues with the 3-step approach (clustering and then predicting), because mixture models are fuzzy clustering and you have to determine the group membership first (based on probability) and then predict, which introduces errors. These models often result in weaker associations. There are some advanced models (bias-corrected), but I don't think there is a conclusion on the best model, yet. Anyway, I think this is the price that I am willing to pay as many other clustering algorithms will suffer from the same issue.
I would vote for the 3-step approach in general unless there is a specific reason for the 1-step model.
- Badges
- Science topic
- Similar topics
- Phytochemistry
- Extracts